Explainable machine-learning predictions for the prevention of hypoxaemia during surgery

Predicting prolonged length of in-hospital stay in patients with non-ST elevation myocardial infarction (NSTEMI) using artificial intelligence

BACKGROUND

Patients presenting with non-ST elevation myocardial infarction (NSTEMI) are typically evaluated using coronary angiography and managed through coronary revascularization. Numerous studies have demonstrated the benefits of expedited discharge following revascularization in this patient population. However, individuals with concomitant heart failure, hemodynamic instability, or arrhythmias often necessitate prolonged hospitalization. Using aortic pressure (AP) wave assessment, we aim to predict a prolonged length of stay (> 4 days, PLoS) in patients with NSTEMI treated with percutaneous coronary intervention (PCI).

METHODS

In this single-center, retrospective cohort study, we included 497 patients with NSTEMI [66.3 ± 12.9 years, 37.6 % (187) females]. We developed a predictive model for PLoS using features primarily extracted from the AP signal recorded throughout PCI. We performed feature selection using recursive feature elimination (RFE) with cross-validation and built a machine learning (ML) model using the CatBoost tree-based classifier. The decision-making process of the ML model was analyzed using SHapley Additive exPlanations (SHAP).

RESULTS

We achieved average accuracy, specificity, sensitivity, precision, and receiver operating characteristic curve area under the curve (AUC) values of 77 %, 78 %, 76 %, 67 %, and 77 %, respectively. Using SHAP, we identified the ejection systolic period, ejection systolic time, the difference between systolic blood pressure and dicrotic notch pressure (DesP), the age modified shock index (mSI_age) and mean arterial pressure (MAP) as the most characteristic features extracted from the AP signal.

CONCLUSIONS

In conclusion, this study demonstrates the potential of using ML and features extracted from the AP signal to predict PLoS in patients with NSTEMI.

Explainable machine learning models enhance prediction of PFAS bioactivity using quantitative molecular surface analysis-derived representation

The extensive use of per- and polyfluoroalkyl substances (PFAS) in industrial and consumer products poses health risks due to their toxicity. Computational toxicology approaches, particularly quantitative structure-activity relationship (QSAR) models are essential for predicting PFAS bioactivity. However, established QSAR models including machine learning-based ones with traditional molecular descriptors such as constitutional, topological, and geometric descriptors, have limited predictive capability and interpretability. Herein, we proposed a novel machine learning approach that leverages quantitative molecular surface analysis (QMSA) of molecular electrostatic potential. Using QMSA descriptors, five machine learning models (e.g., random forest) achieved outstanding performance, with best accuracy of 0.950 ± 0.017, AUC-ROC of 0.938 ± 0.012, F1-score of 0.734 ± 0.024, and MCC of 0.684 ± 0.111 for five targets (tyrosyl-DNA phosphodiesterase 1 in the absence/presence of camptothecin, ATXN2 protein, transcription factor SMAD3, and transcription factor NRF2), which outperform previously reported models. SHAP analyses revealed that estimated density, molecular volume, positive surface area, and nonpolar surface area were the most important descriptors. These descriptors were deeply involved in PFAS binding to target proteins via non-covalent interactions as evidenced by molecular docking and molecular dynamics simulations. Our results demonstrated that QMSA descriptors-based machine learning models are capable of predicting PFAS toxicity with extraordinary performance and interpretability. This study provides a novel machine learning framework for the high-throughput and cost-effective screening of high-risk emerging PFAS in aquatic environments. By identifying the contaminants that should be prioritized for regulation and treatment among the growing number of PFAS, our work aids in water quality monitoring and risk assessment, and guides decision-making in aquatic environmental management. Furthermore, this work enhances our understanding of the molecular mechanisms involved in PFAS bioactivity.

Impact of 2D and 3D factors on urban flooding: Spatial characteristics and interpretable analysis of drivers

Urban flooding poses a serious threat to both the ecological environment and human society. Previous studies identified natural and anthropogenic factors as contributors to urban flooding, but little attention has been paid to the influence of urban horizontal and vertical factors. To address this gap, we conducted a comparative analysis of the patterns in spatial distribution of urban flooding in two megacities in eastern China (Beijing and Guangzhou). We then used Pearson's correlation to investigate the associations between flooding events and multiple influencing factors. Finally, two scenarios were designed to quantify the relative contributions of each driver using the Light Gradient Boosting Machine (LightGBM) and Shapley (SHAP) interpretable models. The results show that: (1) urban flooding points in Guangzhou and Beijing are predominantly clustered in central areas, with mid-rise and high-density buildings presenting the highest flood risk. (2) in the base scenario, Annual precipitation (AP) is the primary influencing factor for urban flooding in both Beijing and Guangzhou. However, in the enhanced scenario, the addition of 2D and 3D (two-dimensional and three-dimensional) metrics shifts the main drivers to factors like Aggregation index (AI), Patch density (PD), and Building density (BD), significantly impacting urban flooding. This study highlights the critical impacts of horizontal and vertical urban structures and layouts, emphasizing the need for comprehensive urban planning and design strategies to effectively mitigate flood risk. It also provides new perspectives on urban flood risk management.

Framingham Risk Score Prediction at 12 Months in the STANDFIRM Randomized Control Trial

BACKGROUND

The STANDFIRM (Shared Team Approach Between Nurses and Doctors for Improved Risk Factor Management; ANZCTR registration ACTRN12608000166370) trial was designed to test the effectiveness of chronic disease care management for modifying the Framingham risk score (FRS) among patients with stroke or transient ischemic attack. The primary outcome of change in FRS was not met. We determine baseline characteristics that predict reduction in FRS at 12 months and whether future FRS is predetermined at baseline.

METHODS AND RESULTS

We used machine learning regression methods to evaluate 35 variables encompassing demographics, risk factors, psychological, social and education status, and laboratory tests. We determine the optimal machine learning and associated tuning parameters from the following: random forest, extreme gradient boosting, category boosting, support vector regression, multilayer perceptron neural network, and K-nearest neighbor. Training (n=404) and test (n=103) data were evenly matched for age, sex, baseline, and 12-month FRS. The optimal model for predicting FRS at 12 months was category boosting (R2=0.712; root mean square error, 7.32). The 5 variables with highest Shapley values for category boosting were baseline FRS (Shapley additive explanation [SHAP], 8.42 of total of 12.12), age (SHAP, 1.58), systolic blood pressure (SHAP, 0.23), male sex (SHAP, 1.05), and London Handicap (SHAP, 0.18). Machine learning methods were poor at determining change in FRS at 12 months (R2<0.22).

CONCLUSIONS

Our findings suggest that change in FRS as an end point in secondary stroke trials may have limited value as it is largely determined at baseline. In this cohort, category boosting was the optimal method to predict future FRS but not change in FRS.

Evaluating the SWIFT algorithm's efficacy in predicting hypoxemia across multiple critical care datasets

BACKGROUND

Machine learning models to predict hypoxia in patients could improve timely interventions. Due to the diversity and limited generalizability of approaches, external validation is required.

OBJECTIVE

This study aimed to validate the generalizability of SpO2 Waveform ICU Forecasting Technique (SWIFT), an LSTM algorithm for predicting SpO2 5 and 30 min in advance, on two external datasets.

METHODS

We trained the SWIFT model on eICU Collaborative Research Database (eICU-CRD) and validated it on Medical Information Mart for Intensive Care IV (MIMIC-IV) and Amsterdam University Medical Centers Database (UMCdb) data. We evaluated SWIFT-5 and SWIFT-30 for ventilated and non-ventilated populations.

RESULTS

The sampling procedure resulted in substantial population size reduction for MIMIC-IV and UMCdb data due to differences in SpO2 measurement frequency. SWIFT performed well on eICU-CRD data but showed reduced performance on MIMIC-IV data, particularly for SWIFT-30. UMCdb validation demonstrated promise, with comparable performance to eICU-CRD for ventilated patients. All datasets exhibited high specificity and NPV, critical for gaining trust in alarms in clinical applications.

CONCLUSIONS

The study highlights challenges in generalizing prediction models across diverse ICU populations, emphasizing need for external validation. Further research should focus on improving model adaptability and interpretability, considering the practical application in clinical settings.

Development and validation of clinical-radiomics deep learning model based on MRI for endometrial cancer molecular subtypes classification

OBJECTIVES

This study aimed to develop and validate a clinical-radiomics deep learning (DL) model based on MRI for endometrial cancer (EC) molecular subtypes classification.

METHODS

This multicenter retrospective study included EC patients undergoing surgery, MRI, and molecular pathology diagnosis across three institutions from January 2020 to March 2024. Patients were divided into training, internal, and external validation cohorts. A total of 386 handcrafted radiomics features were extracted from each MR sequence, and MoCo-v2 was employed for contrastive self-supervised learning to extract 2048 DL features per patient. Feature selection integrated selected features into 12 machine learning methods. Model performance was evaluated with the AUC.

RESULTS

A total of 526 patients were included (mean age, 55.01 ± 11.07). The radiomics model and clinical model demonstrated comparable performance across the internal and external validation cohorts, with macro-average AUCs of 0.70 vs 0.69 and 0.70 vs 0.67 (p = 0.51), respectively. The radiomics DL model, compared to the radiomics model, improved AUCs for POLEmut (0.68 vs 0.79), NSMP (0.71 vs 0.74), and p53abn (0.76 vs 0.78) in the internal validation (p = 0.08). The clinical-radiomics DL Model outperformed both the clinical model and radiomics DL model (macro-average AUC = 0.79 vs 0.69 and 0.73, in the internal validation [p = 0.02], 0.74 vs 0.67 and 0.69 in the external validation [p = 0.04]).

CONCLUSIONS

The clinical-radiomics DL model based on MRI effectively distinguished EC molecular subtypes and demonstrated strong potential, with robust validation across multiple centers. Future research should explore larger datasets to further uncover DL's potential.

CRITICAL RELEVANCE STATEMENT

Our clinical-radiomics DL model based on MRI has the potential to distinguish EC molecular subtypes. This insight aids in guiding clinicians in tailoring individualized treatments for EC patients.

KEY POINTS

Accurate classification of EC molecular subtypes is crucial for prognostic risk assessment. The clinical-radiomics DL model outperformed both the clinical model and the radiomics DL model. The MRI features exhibited better diagnostic performance for POLEmut and p53abn.

Effects of neonicotinoid pesticide exposure in the first trimester on gestational diabetes mellitus based on interpretable machine learning

BACKGROUND

Gestational diabetes mellitus (GDM) is one of the most common pregnancy complications and seriously threatens the health of mothers and offspring. Neonicotinoids (NEOs) is a new class of pesticide and widely used worldwide. Prenatal NEOs exposure had negative effects on fetal growth, but the potential effect of NEOs exposure on pregnancy complications remain unclear.

OBJECTIVES

To examine the individual and jointed effects of serum neonicotinoids (NEOs) pesticide exposure on gestational diabetes mellitus (GDM), and explore the application of NEOs exposure levels as predictor of GDM.

METHODS

We conducted a prospective cohort study based on Guangxi Zhuang Birth Cohort, China. A total of 1450 mather-infant pairs were included from 2015 to 2019. Ten NEOs were measured by UPLC-MS. Maternal serum samples were collected during gestational age 0-12 weeks. Individual and jointed effects of NEOs on GDM were assessed through binomial regressions, Bayesian Kernel Machine Regression and quantile g-computation. Prediction of GDM using XGboost machine learning and SHapley Additive exPlanations (SHAP).

RESULTS

A total of 122 (8.4%) mothers were diagnosed with GDM. In the individual exposure models, sulfoxaflor and thiamethoxam exposure in the first trimester significantly increased the risk of GDM (OR = 1.48, 95%CI: 1.21, 1.82; OR = 1.42, 95%CI: 1.14, 1.78). Moreover, GDM risk increased significantly with NEOs mixture concentration was above 75th percentile, compared with the 50th percentile. Sulfoxaflor and thiamethoxam as the main positive contributing factors in NEOs mixture to increase the GDM with a weight of 29.3% and 27.6%, respectively. Furthermore, sulfoxaflor and thiamethoxam were the most important contributing factors for predicting GDM after combining traditional risk factors in machine learning model, with predicted contribution values of 0.79 and 0.46, respectively.

CONCLUSION

Our findings suggested that elevated maternal serum sulfoxaflor, thiamethoxam and NEOs mixture were positively associated with GDM, and sulfoxaflor, thiamethoxam were the important contributing factors for predicting GDM.

Distinguishing critical microbial community shifts from normal temporal variability in human and environmental ecosystems

Differentiating significant microbial community changes from normal fluctuations is vital for understanding microbial dynamics in human and environmental ecosystems. This knowledge could enable early warning systems to monitor critical changes affecting human or environmental health. We applied 16S rRNA gene sequencing and time-series analysis to model bacterial abundance trajectories in human gut and wastewater microbiomes. We evaluated various model architectures using datasets from two human studies and five wastewater settings. Long short-term memory (LSTM) models consistently outperformed other models in predicting bacterial abundances and detecting outliers, as measured by multiple metrics. Prediction intervals for each genus allowed us to identify significant changes and signaling shifts in community states. This study proposes a machine learning model capable of monitoring microbial communities and providing insights into their responses to internal and external factors in medical and environmental settings.

Quantifying role of source variations on PM2.5-bound toxic components under climate change: Measurement at multiple sites during 2018-2022 in a Chinese megacity

Understanding the response of PM2.5-bound toxic components to source variations under climate change is crucial for public health protection. However, the lack of long-term and multi-site observational data of toxic components limits such efforts. Here, we conducted a five-year PM2.5 measurement (2018-2022) at 10 sites across a Chinese megacity, analyzing 15 polycyclic aromatic hydrocarbons (PAHs), 6 organophosphate esters (OPEs), and 9 potentially toxic elements (PTEs). Using explainable machine learning, we found that source variations from particle matter mass reduction under climate change can impact PM2.5-bound toxic components. Meteorological factors like extreme heat days and max temperature impact most toxic components, while geographic, socioeconomic, and anthropogenic factors mainly affect PTEs, especially Cu. We also designed 10 extreme heat and source variation scenarios to predict the response of toxic components. When comparing scenario 2-1 (source variation without temperature change) with scenario 2-2 and 2-3 (the same source variation but higher temperatures), many PM2.5-bound organics and As show higher reduction rates under climate change, highlighting the need to focus more on gas-phase organics and products of atmospheric process. Benzo[b]fluoranthene (BbF) is most sensitive to traffic source reductions, and Cu, Mn, Zn and Fe are more sensitive to industrial source reductions.

Predicting onward care needs at admission to reduce discharge delay using explainable machine learning

Early identification of patients who require onward referral to social care can prevent delays to discharge from hospital. We introduce an explainable machine learning (ML) model to identify potential social care needs at the first point of admission. This model was trained using routinely collected data on patient admissions, hospital spells and discharge at a large tertiary hospital in the UK between 2017 and 2023. The model performance (one-vs-rest AUROC = 0.915 [0.907 0.924] (95% confidence interval), is comparable to clinician's predictions of discharge care needs, despite working with only a subset of the information available to the clinician. We find that ML and clinicians perform better for identifying different types of care needs, highlighting the added value of a potential system supporting decision making. We also demonstrate the ability for ML to provide automated initial discharge need assessments, in the instance where initial clinical assessment is delayed and provide reasoning for the decision. Finally, we demonstrate that combining clinician and machine predictions, in a hybrid model, provides even more accurate early predictions of onward social care requirements (OVR AUROC = 0.936 [0.928 0.943]) and demonstrates the potential for human-in-the-loop decision support systems in clinical practice.

Navigating the Fragrance Space Using Graph Generative Models and Predicting Odors

We explore a suite of generative modeling techniques to efficiently navigate and explore the complex landscapes of odor and the broader chemical space. Unlike traditional approaches, we not only generate molecules but also predict the odor likeliness with a ROC AUC score of 0.97 and assign probable odor labels. We correlate odor likeliness with physicochemical features of molecules using machine learning techniques and leverage SHAP (SHapley Additive exPlanations) to demonstrate the interpretability of the function. The whole process involves four key stages: molecule generation, stringent sanitization checks for molecular validity, fragrance likeliness screening, and odor prediction of the generated molecules. By making our code and trained models publicly accessible, we aim to facilitate the broader adoption of our research across applications in fragrance discovery and olfactory research.

Integrating SHAP analysis with machine learning to predict postpartum hemorrhage in vaginal births

OBJECTIVE

This study aimed to develop a machine learning (ML) model integrated with SHapley Additive exPlanations (SHAP) analysis to predict postpartum hemorrhage (PPH) following vaginal deliveries, offering a potential tool for personalized risk assessment and prevention in clinical settings.

METHODS

We conducted a retrospective multicenter cohort study in Northeast China, including women who had vaginal deliveries at three tertiary hospitals from September 2018 to December 2023. Data were extracted from electronic medical records. The dataset was split into a training set (70%) and an internal validation set (30%) to prevent overfitting. External validation was performed on a separate dataset. Several evaluation metrics, including the area under the receiver operating characteristic curve (AUC), were used to compare prediction performance. Features were ranked using SHAP, and the final model was explained.

RESULTS

The XGBoost model demonstrated superior predictive accuracy for PPH, with an AUC of 0.997 in the training set. SHAP value-based feature selection identified 15 key features contributing to the model's predictive power. SHAP dependence and summary plots provided intuitive insights into each feature's contribution, enabling the identification of anomalies. The final model maintained high predictive power, with an AUC of 0.894 in internal validation and 0.880 in external validation.

CONCLUSION

This study successfully developed an interpretable ML model that predicts PPH with high accuracy. Future studies with larger and more diverse datasets are necessary to further validate and refine the model, particularly to assess its generalizability across different populations and healthcare settings.

Resistive Switching Random-Access Memory (RRAM): Applications and Requirements for Memory and Computing

In the information age, novel hardware solutions are urgently needed to efficiently store and process increasing amounts of data. In this scenario, memory devices must evolve significantly to provide the necessary bit capacity, performance, and energy efficiency needed in computation. In particular, novel computing paradigms have emerged to minimize data movement, which is known to contribute the largest amount of energy consumption in conventional computing systems based on the von Neumann architecture. In-memory computing (IMC) provides a means to compute within data with minimum data movement and excellent energy efficiency and performance. To meet these goals, resistive-switching random-access memory (RRAM) appears to be an ideal candidate thanks to its excellent scalability and nonvolatile storage. However, circuit implementations of modern artificial intelligence (AI) models require highly specialized device properties that need careful RRAM device engineering. This work addresses the RRAM concept from materials, device, circuit, and application viewpoints, focusing on the physical device properties and the requirements for storage and computing applications. Memory applications, such as embedded nonvolatile memory (eNVM) in novel microcontroller units (MCUs) and storage class memory (SCM), are highlighted. Applications in IMC, such as hardware accelerators of neural networks, data query, and algebra functions, are illustrated by referring to the reported demonstrators with RRAM technology, evidencing the remaining challenges for the development of a low-power, sustainable AI.

The Importance of Atomic Charges for Predicting Site-Selective Ir-, Ru-, and Rh-Catalyzed C-H Borylations

A supervised machine learning model has been developed that allows for the prediction of site selectivity in late-stage C-H borylations. Model development was accomplished using literature data for the site-selective (≥95%) C-H borylation of 189 unique arene, heteroarene, and aliphatic substrates that feature a total of 971 possible sp2 or sp3 C-H borylation sites. The reported experimental data was supplemented with additional chemoinformatic descriptors, computed atomic charges at the C-H borylation sites, and data from parameterization of catalytically active tris-boryl complexes resulting from the combination of seven different Ir-, Ru-, and Rh-based precatalysts with eight different ligands. Of the over 1600 parameters investigated, the computed atomic charges (e.g., Hirshfeld, ChelpG, and Mulliken charges) on the hydrogen and carbon atoms at the site of borylation were identified as the most important features that allow for the successful prediction of whether a particular C-H bond will undergo a site-selective borylation. The overall accuracy of the developed model was 88.9% ± 2.5% with precision, recall, and F1 scores of 92-95% for the nonborylating sites and 65-75% for the sites of borylation. The model was demonstrated to be generalizable to molecules outside of the training/test sets with an additional validation set of 12 electronically and structurally diverse systems.

Visualizing functional network connectivity differences using an explainable machine-learning method

Objective. Functional network connectivity (FNC) estimated from resting-state functional magnetic resonance imaging showed great information about the neural mechanism in different brain disorders. But previous research has mainly focused on standard statistical learning approaches to find FNC features separating patients from control. While machine learning models can improve classification accuracy, they often lack interpretability, making it difficult to understand how they arrive at their decisions.Approach. Explainable machine learning helps address this issue by identifying which features contribute most to the model's predictions. In this study, we introduce a novel framework leveraging SHapley Additive exPlanations (SHAPs) to identify crucial FNC features distinguishing between two distinct population classes.Main results. Initially, we validate our approach using synthetic data. Subsequently, applying our framework, we ascertain FNC biomarkers distinguishing between, controls and schizophrenia (SZ) patients with accuracy of 81.04% as well as middle aged adults and old aged adults with accuracy 71.38%, respectively, employing random forest, XGBoost, and CATBoost models.Significance. Our analysis underscores the pivotal role of the cognitive control network (CCN), subcortical network (SCN), and somatomotor network in discerning individuals with SZ from controls. In addition, our platform found CCN and SCN as the most important networks separating young adults from older.

Machine learning prediction models for multidrug-resistant organism infections in ICU ventilator-associated pneumonia patients: Analysis using the MIMIC-IV database

OBJECTIVE

This study aims to construct and compare four machine learning models using the MIMIC-IV database to identify high-risk factors for multidrug-resistant organism (MDRO) infection in Ventilator-associated pneumonia (VAP) patients.

METHODS

The study included 972 VAP patients from the MIMIC-IV database. Data encompassing demographic information, vital signs, laboratory results, and other relevant variables were collected. The class imbalance issue was addressed using the Synthetic Minority Over-sampling Technique (SMOTE). The dataset was randomly split into training and testing sets (8:2). LASSO regression and feature importance scores were used for feature selection. Clinical prediction models were built using logistic regression, XGBoost, random forest and gradient boosting machine. The performance of the models was evaluated through receiver operating characteristic(ROC) curve analysis.Model calibration was assessed using calibration curves and Brier scores. The effectiveness was evaluated through Decision Curve Analysis (DCA). SHAP was utilized for model interpretation.

RESULTS

Among 972 patients, 824 were non-MDROs-VAP and 128 were MDROs-VAP. Comparative analysis revealed statistically significant differences in various clinical parameters. XGBoost exhibited the best predictive performance, incorporating 20 features with an AUC of 0.831 (95 % CI: 0.785-0.877) on the test set. Calibration curves demonstrated robust consistency, corroborated by Decision Curve Analysis (DCA) affirming the clinical utility. SHAP analysis identified the most important features: red cell distribution width, duration of mechanical ventilation, anion gap, basophil percentage, and neutrophil percentage.

CONCLUSION

This study established and compared four machine learning models for MDROs infections in VAP patients. XGBoost was identified as the optimal predictor, and SHAP values provided insights into 20 independent risk factors, confirming its excellent predictive value.

IMPLICATIONS FOR CLINICAL PRACTICE

VAP is a common infection in ICU patients with a heightened risk of MDRO and increased mortality. The recognition of high bias in existing models calls for future research to employ rigorous methodologies and robust data sources, aiming to develop and validate more accurate and clinically applicable predictive models for MDROs infections in VAP patients.

Explainable Machine Learning Models for Prediction of Surgical Site Infection After Posterior Lumbar Fusion Surgery Based on Shapley Additive Explanations

OBJECTIVE

This study aims to develop machine learning (ML) models combined with an explainable method for the prediction of surgical site infection (SSI) after posterior lumbar fusion surgery.

METHODS

In this retrospective, single-center study, a total of 1016 consecutive patients who underwent posterior lumbar fusion surgery were included. A comprehensive dataset was established, encompassing demographic variables, comorbidities, preoperative evaluation, details related to diagnosed lumbar disease, preoperative laboratory tests, surgical specifics, and postoperative factors. Utilizing this dataset, 6nullML models were developed to predict the occurrence of SSI. Performance evaluation of the models on the testing set involved several metrics, including the receiver operating characteristic curve, the area under the receiver operating characteristic curve, accuracy, recall, F1 score, and precision. The Shapley Additive Explanations (SHAP) method was employed to generate interpretable predictions, enabling a comprehensive assessment of SSI risk and providing individualized interpretations of the model results.

RESULTS

Among the 1016 retrospective cases included in the study, 36 (3.54%) experienced SSI. Out of the six models examined, the Extreme Gradient Boost model demonstrated the highest discriminatory performance on the testing set, achieving the following metrics: precision (0.9000), recall (0.8182), accuracy (0.9902), F1 score (0.8571), and area under the receiver operating characteristic curve (0.9447). By utilizing the SHAP method, several important predictors of SSI were identified, including the duration of indwelling jugular vein catheter, blood urea nitrogen levels, total protein levels, sustained fever, creatinine levels, triglycerides levels, monocyte count, diabetes mellitus, drainage time, white blood cell count, cerebral infarction, estimated blood loss, prealbumin levels, Prognostic Nutritional Index, low back pain, posterior fusion score, and osteoporosis.

CONCLUSIONS

ML-based prediction tools can accurately assess the risk of SSI after posterior lumbar fusion surgery. Additionally, ML combined with SHAP could provide a clear interpretation of individualized risk prediction and give physicians an intuitive comprehension of the effects of the model's essential features.

Development and Validation of Multiparametric MRI-based Interpretable Deep Learning Radiomics Fusion Model for Predicting Lymph Node Metastasis and Prognosis in Rectal Cancer: A Two-center Study

RATIONALE AND OBJECTIVES

To develop interpretable machine learning models that utilize deep learning (DL) and radiomics based on multiparametric Magnetic resonance imaging (MRI) to predict preoperative lymph node (LN) metastasis in rectal cancer.

MATERIALS AND METHODS

This retrospective study involved 286 cancer patients confirmed by histopathology from center 1 (Training set) and 66 patients from center 2 (External test set). Radiomics features were extracted from T2-weighted imaging (T2WI) and diffusion-weighted imaging (DWI) sequences, whereas DL features were obtained using four models: MobileNet-V3-large, Inception-V3, ResNet50, and VGG16. These DL radiomics (DLR) features were then combined to construct a machine learning model. The Shapley additive interpretation (SHAP) tool was utilized to investigate the interpretability of the model. We evaluated and compared the diagnostic performance of senior and junior radiologists, with and without the aid of the optimal DLR model. Kaplan-Meier survival curve was used to analyze the prognosis of patients.

RESULTS

The DLR model outperforms individual DL models and the radiomics model. The MobileNet-V3-large combination radiomics signature demonstrated the best performance, achieving an AUC of 0.878 on the Training set and 0.752 on the External test set. Compared to the traditional radiomics model, the AUC for the Training set increased by 0.094 and by 0.051 for the External test set. This model facilitated improved diagnostic performance among both junior and senior radiologists. Specifically, the AUC values for junior and senior radiologists increased by 0.162 and 0.232, respectively, on the Training set; and by 0.096 and 0.113, respectively, on the External test set. The DLR model demonstrated strong performance in risk stratification for disease-free survival.

CONCLUSION

The DLR model developed from multiparametric MRI can effectively distinguish cancer LN metastasis and enhance radiologists' diagnostic performance.

Machine learning and complex network analysis of drug effects on neuronal microelectrode biosensor data

Biosensors, such as microelectrode arrays that record in vitro neuronal activity, provide powerful platforms for studying neuroactive substances. This study presents a machine learning workflow to analyze drug-induced changes in neuronal biosensor data using complex network measures from graph theory. Microelectrode array recordings of neuronal networks exposed to bicuculline, a GABA[Formula: see text] receptor antagonist known to induce hypersynchrony, demonstrated the workflow's ability to detect and characterize pharmacological effects. The workflow integrates network-based features with synchrony, optimizing preprocessing parameters, including spike train bin sizes, segmentation window sizes, and correlation methods. It achieved high classification accuracy (AUC up to 90%) and used Shapley Additive Explanations to interpret feature importance rankings. Significant reductions in network complexity and segregation, hallmarks of epileptiform activity induced by bicuculline, were revealed. While bicuculline's effects are well established, this framework is designed to be broadly applicable for detecting both strong and subtle network alterations induced by neuroactive compounds. The results demonstrate the potential of this methodology for advancing biosensor applications in neuropharmacology and drug discovery.

Global trends in artificial intelligence research in anesthesia from 2000 to 2023: a bibliometric analysis

BACKGROUND

Interest in artificial intelligence (AI) research in anesthesia is growing rapidly. However, there is a lack of bibliometric analysis to measure and analyze global scientific publications in this field. The aim of this study was to identify the hotspots and trends in AI research in anesthesia through bibliometric analysis.

METHODS

English articles and reviews published from 2000 to 2023 were retrieved from the Web of Science Core Collection (WoSCC) database. The extracted data were summarized and analyzed using Microsoft Excel, and bibliometric analysis were conducted with VOSviewer software.

RESULTS

AI research literature in anesthesia has exhibited rapid growth in recent years. The United States leads in the number of publications and citations, with Stanford University as the most prolific institution. Hyung-Chul Lee is the author with the highest number of publications. The journal Anesthesiology is highly recognized and authoritative in this field. Recent keywords include "musculoskeletal pain", "precision medicine", "stratification", "images", "mean arterial pressure", " enhanced recovery after surgery", "frailty", "telehealth", "postoperative delirium" and "postoperative mortality" indicating hot topics in AI research in anesthesia.

CONCLUSIONS

Publications on AI research in the field of anesthesia have experienced rapid growth over the past two decades and are likely to continue increasing. Research areas such as depth of anesthesia (DOA) and drug infusion (including electroencephalography and deep learning), perioperative risk assessment and prediction (covering mean arterial pressure, frailty, postoperative delirium, and mortality), image classification and recognition (for applications such as ultrasound-guided nerve blocks, vascular access, and difficult airway assessment), and perioperative pain management (particularly musculoskeletal pain) have garnered significant attention. Additionally, topics such as precision medicine, enhanced recovery after surgery, and telehealth are emerging as new hotspots and future directions in this field.

A Review of Non-Cardiac Complications of General Anesthesia: The Current State of Knowledge

General anesthesia, despite the constant development of anesthesiology, still carries certain risks. To provide safe anesthesia, it is crucial to properly qualify patients and to react in an appropriate manner when problems occur. It is therefore essential to have knowledge of risk factors, pathophysiology, symptoms, and management patterns regarding complications. This review comprehensively describes respiratory complications such as airway spasm, conditions leading to intraoperative hypoxemia, postoperative pulmonary complications (PPC), and complications of cross airway compromise, from aspects including respiratory complications and mechanical injuries. Moreover, events characteristic of this type of anesthesia, such as anaphylaxis, postoperative nausea and vomiting (PONV), neurological complications, accidental awakening during general anesthesia (AAGA), hypothermia, and malignant hyperthermia (MH), have been included. Each complication is elaborated on in terms of risk groups and factors, symptoms, and prevention and treatment options, taking into account the interrelationship of particular conditions. Although that issue is well reported in the literature, this review, in addition to a comprehensive summary of the most important non-cardiovascular and hemodynamic complications, takes into account the latest findings on methods of prevention, diagnosis, and intraoperative monitoring. The article combines a comprehensive compilation of basic information on the most significant complications, including their diagnosis and methods of intervention, along with consideration of the latest scientific developments and indication of future research directions. This review is based on the most recent articles possible, published between 2006 and 2024.

Constructing a screening model to identify patients at high risk of hospital-acquired influenza on admission to hospital

Objective

To develop a machine learning (ML)-based admission screening model for hospital-acquired (HA) influenza using routinely available data to support early clinical intervention.

Methods

The study focused on hospitalized patients from January 2021 to May 2024. The case group consisted of patients with HA influenza, while the control group comprised non-HA influenza patients admitted to the same ward in the HA influenza unit within 2 weeks. The 953 subjects were divided into the training set and the validation set in a 7:3 ratio. Feature screening was performed using least absolute shrinkage and selection operator (LASSO) and the Boruta algorithm. Subsequently eight ML algorithms were applied to analyze and identify the optimal model using a 5-fold cross-validation methodology. And the area under the curve (AUC), area under the precision-recall curve (AP), F1 score, calibration curve and decision curve analysis (DCA) were applied to comprehensively assess the predictive effectiveness of the selected models. Feature factors were selected and feature importance's were assessed using SHapley's additive interpretation (SHAP). Furthermore, an interactive web-based platform was additionally developed to visualize and demonstrate the predictive model.

Results

Age, pneumonia on admission, Chronic renal failure, Malignant tumor, hypoproteinemia, glucocorticoid use, admission to ICU, lymphopenia, BMI were identified as key variables. For the eight ML algorithms, ROC values ranging from 0.548 to 0.812 were observed in the validation set. A comprehensive analysis showed that the XGBoost model predicted the highest accuracy (AUC: 0.812) with an F1 score of 0.590 and the highest A p value (0.655). Evaluating the optimal model, the AUC values were 0.995, 0.826, and 0.781 for the training, validation and test sets. The XGBoost model showed strong robust. SHapley's additive interpretation (SHAP) was utilized to analyze the contribution of explanatory variables to the model and their correlation with HA influenza. In addition, we developed a practical online prediction tool to calculate the risk of HA influenza occurrence.

Conclusion

Based on the routine data, the XGBoost model demonstrated excellent calibration among all ML algorithms and accurately predicted the risk of HA influenza, thereby serving as an effective tool for early screening of HA influenza.

Analysis of risk factors for hypoxemia in PACU for patients undergoing thoracoscopic lung cancer resection based on logistic regression model

OBJECTIVE

This study aims to identify risk factors of hypoxemia in patients undergoin thoracoscopic lung surgery during their stay in the post-anesthesia care unit (PACU). Hypoxemia was defined as any instance of SpO₂ ≤90% lasting for more than one minute during the PACU stay.

METHODS

We conducted a prospective research involving 398 patients who underwent elective thoracoscopic lung surgery in West China Hospital, Sichuan University, from April to July 2024. Patients were classified into hypoxemia and non-hypoxemia groups based on the presence of hypoxemia in the PACU. We compared clinical data between the two groups to identify factors influencing hypoxemia. Variables with statistical significance (P < 0.05) in univariate analysis were included in logistic regression to identify independent risk factors for hypoxemia.

RESULTS

Among the 398 patients studied, 149 (37.4%) experienced hypoxemia. Univariate analysis indicated significant differences in age, BMI, height, ASA classification, hypertension, diabetes, lung function test with Forced Expiratory Volume at 1 s / Forced Vital Capacity (FEV1/FVC), and awakening time between the groups. Logistic regression revealed that age, BMI, ASA classification, hypertension, diabetes, and awakening time were independent risk factors for hypoxemia during anesthesia recovery, while preoperative SpO2 upon entering operating room (OR = 0.882, 95% CI: 0.783-0.993, P = 0.038) was identified as a protective factor.

CONCLUSION

Age, BMI, ASA classification, and preoperative conditions such as hypertension and diabetes are found to contribute to an increased incidence of hypoxemia in PACU following thoracoscopic lung surgery. Emphasizing preoperative lung function assessments and enhanced monitoring may also facilitate timely interventions, thereby improving post-anesthesia recovery and patient outcomes.

A genotype-phenotype transformer to assess and explain polygenic risk

Genome-wide association studies have linked millions of genetic variants to biomedical phenotypes, but their utility has been limited by lack of mechanistic understanding and widespread epistatic interactions. Recently, Transformer models have emerged as a powerful machine learning architecture with potential to address these and other challenges. Accordingly, here we introduce the Genotype-to-Phenotype Transformer (G2PT), a framework for modeling hierarchical information flow among variants, genes, multigenic systems, and phenotypes. As proof-of-concept, we use G2PT to model the genetics of TG/HDL (triglycerides to high-density lipoprotein cholesterol), an indicator of metabolic health. G2PT predicts this trait via attention to 1,395 variants underlying at least 20 systems, including immune response and cholesterol transport, with accuracy exceeding state-of-the-art. It implicates 40 epistatic interactions, including epistasis between APOA4 and CETP in phospholipid transfer, a target pathway for cholesterol modification. This work positions hierarchical graph transformers as a next-generation approach to polygenic risk.

Reinforcement learning for healthcare operations management: methodological framework, recent developments, and future research directions

With the advancement in computing power and data science techniques, reinforcement learning (RL) has emerged as a powerful tool for decision-making problems in complex systems. In recent years, the research on RL for healthcare operations has grown rapidly. Especially during the COVID-19 pandemic, RL has played a critical role in optimizing decisions with greater degrees of uncertainty. RL for healthcare applications has been an exciting topic across multiple disciplines, including operations research, operations management, healthcare systems engineering, and data science. This review paper first provides a tutorial on the overall framework of RL, including its key components, training models, and approximators. Then, we present the recent advances of RL in the domain of healthcare operations management (HOM) and analyze the current trends. Our paper concludes by presenting existing challenges and future directions for RL in HOM.

Explainable Deep Learning to Predict Kelp Geographical Origin from Volatile Organic Compound Analysis

In addition to its flavor and nutritional value, the origin of kelp has become a crucial factor influencing consumer choices. Nevertheless, research on kelp's origin traceability by volatile organic compound (VOC) analysis is lacking, and the application of deep learning in this field remains scarce due to its black-box nature. To address this gap, we attempted to identify the origin of kelp by analyzing its VOCs in conjunction with explainable deep learning. In this work, we identified 115 distinct VOCs in kelp samples using gas chromatography coupled with ion mobility spectroscopy (GC-IMS), of which 68 categories were discernible. Consequently, we developed a comprehensible one-dimensional convolutional neural network (1D-CNN) model that incorporated 107 VOCs exhibiting significant regional disparities (p < 0.05). The model successfully discerns the origin of kelp, achieving perfect metrics across accuracy (100%), precision (100%), recall (100%), F1 score (100%), and AUC (1.0). SHapley Additive exPlanations (SHAP) analysis highlighted the impact of features such as 1-Octen-3-ol-M, (+)-limonene, allyl sulfide-D, 1-hydroxy-2-propanone-D, and (E)-2-hexen-1-al-M on the model output. This research provides deeper insights into how critical product features correlate with specific geographic information, which in turn boosts consumer trust and promotes practical utilization in actual settings.

An interpretable radiomics-based machine learning model for predicting reverse left ventricular remodeling in STEMI patients using late gadolinium enhancement of myocardial scar

OBJECTIVES

To evaluate the added value of the late gadolinium enhancement (LGE)-scar radiomics features in predicting reverse left ventricular remodeling (r-LVR) in ST-segment elevation myocardial infarction (STEMI) patients using machine learning (ML).

MATERIALS AND METHODS

This retrospective study included 105 STEMI patients who underwent CMR within 7 days and 5 months post-percutaneous coronary intervention (PCI) on 1.5-T or 3.0-T MRI scanners (January 2014-2023). Radiomics features from LGE scar images and routine CMR markers were analyzed using a LightGBM model enhanced by Shapley Additive exPlanations (SHAP) for interpretability. Patients were divided into training (80) and test (25) sets. Three predictive models were developed: traditional CMR, LGE-scar radiomics, and a combined model integrating both. Model performance was assessed using ROC curves and AUC analysis.

RESULTS

In the training set, the traditional CMR model achieved an AUC of 0.745 (95% CI: 0.62-0.86), the LGE-scar radiomics model had an AUC of 0.712 (95% CI: 0.58-0.83), and the combined model showed the highest AUC of 0.754 (95% CI: 0.63-0.86). In the test set, the traditional CMR model's AUC decreased to 0.656 (95% CI: 0.42-0.88), while the LGE-scar radiomics model improved to 0.818 (95% CI: 0.59-1.00). The combined model achieved the highest AUC of 0.890 (95% CI: 0.75-1.00). SHAP analysis highlighted significant predictors such as infarct percentage of LV mass and wavelet-transformed texture features.

CONCLUSION

Integrating LGE scar radiomics features with traditional CMR parameters in a LightGBM model enhances predictive accuracy for r-LVR in STEMI patients, potentially improving patient stratification and treatment personalization.

KEY POINTS

Question Predicting r-LVR in STEMI patients remains challenging due to limitations in current imaging approaches. Findings Integrating LGE-scar radiomics and cardiac magnetic resonance markers in the LightGBM model significantly improves prediction accuracy for r-LVR. Clinical relevance This interpretable ML model enhances r-LVR prediction, supporting patient stratification and optimizing treatment strategies to improve patient outcomes.

Arterial phase CT radiomics for non-invasive prediction of Ki-67 proliferation index in pancreatic solid pseudopapillary neoplasms

BACKGROUND

This study aimed to preoperatively predict Ki-67 proliferation levels in patients with pancreatic solid pseudopapillary neoplasm (pSPN) using radiomics features extracted from arterial phase helical CT images.

METHODS

We retrospectively analyzed 92 patients (Ningbo Medical Center Lihuili Hospital: n = 64, Taizhou Central Hospital: n = 28) with pathologically confirmed pSPN from June 2015 to June 2023. Ki-67 positivity > 3% was considered high. Radiomics features were extracted using PyRadiomics, with patients from training cohort (n = 64) and validation cohort (n = 28). A radiomics signature was constructed, and a CT radiomics score (CTscore) was calculated. Deep learning models were employed for prediction, with early stopping to prevent overfitting.

RESULTS

Seven key radiomics features were selected via LASSO regression with cross-validation. The deep learning model demonstrated improved accuracy with demographics and CTscore, with key features such as Morphology and CTscore contributing significantly to predictive accuracy. The best-performing models, including GBM and deep learning algorithms, achieved high predictive performance with an AUC of up to 0.946 in the training cohort.

CONCLUSIONS

We developed a robust deep learning-based radiomics model using arterial phase CT images to predict Ki-67 levels in pSPN patients, identifying CTscore and Morphology as key predictors. This non-invasive approach has potential utility in guiding personalized preoperative treatment strategies.

CLINICAL TRIAL NUMBER

Not applicable.

The association between low-concentration heavy metal exposure and chronic kidney disease risk through α-klotho

Although the association between pollution exposure and chronic kidney disease (CKD) has been explored, previous studies have focused on specific effects observed via in vitro or animal experiments. We first conducted a priority screening of pollutants for population CKD risk by using machine learning approaches. We then used the National Health and Nutrition Examination Survey (NHANES) 2007-2016 data from 2415 adults aged 40 years and over to study the joint effects of low-concentration metal exposure and the mediating effects of α-klotho by using Bayesian kernel machine regression (BKMR) and mediation analyses. Priority screening revealed that cadmium (Cd), mercury (Hg), lead (Pb), and thallium (Tl) were associated with the highest risk of developing CKD. The BKMR model revealed a negative joint effect of mixed-metal exposure on CKD risk. Tl presented the highest posterior inclusion probability (PIP) of 1.0000, followed by Pb, with a PIP of 0.6080. Significant mediating effects of α-klotho on Hg-CKD associations were observed. Mendelian randomization demonstrated that a high level of α-klotho is associated with a decreased risk of developing CKD. This is the first study to reveal the risk prioritization of various pollutants in CKD patients, as well as the coexposure effects of metals. Our study also provides insight into the potential mechanisms underlying the association between metal exposure and CKD risk.

A unified ensemble soil moisture dataset across the continental United States

A unified ensemble soil moisture (SM) package has been developed over the Continental United States (CONUS). The data package includes 19 products from land surface models, remote sensing, reanalysis, and machine learning models. All datasets are unified to a 0.25-degree and monthly spatiotemporal resolution, providing a comprehensive view of surface SM dynamics. The statistical analysis of the datasets leverages the Koppen-Geiger Climate Classification to explore surface SM's spatiotemporal variabilities. The extracted SM characteristics highlight distinct patterns, with the western CONUS showing larger coefficient of variation values and the eastern CONUS exhibiting higher SM values. Remote sensing datasets tend to be drier, while reanalysis products present wetter conditions. In-situ SM observations serve as the basis for wavelet power spectrum analyses to explain discrepancies in temporal scales across datasets facilitating daily SM records. This study provides a comprehensive soil moisture data package and an analysis framework that can be used for Earth system model evaluations and uncertainty quantification, quantifying drought impacts and land-atmosphere interactions and making recommendations for drought response planning.

Development of a multi-module data-driven integrated framework for identifying drivers of atmospheric particulate nitrate and reduction emissions: An application in an industrial city, China

Atmospheric particulate nitrate (pNO3-), a crucial component of fine particulate matter, significantly contributes to haze pollution. The formation of pNO3- is driven by multiple factors including meteorology, emissions, and atmospheric chemistry. Understanding the key drivers of pNO3- formation and developing an accurate and physically meaningful method for the timely assessment of the direct causes of pNO3- pollution are essential. In this study, we propose a multi-module data-driven integrated framework that incorporates and improves four distinct machine learning modules. This framework enhances the physical interpretability of the statistical outcomes of the driving factors of pNO3-, quantifies the impacts of multiple factors on pNO3-, and reveals emission reduction trends. Our findings show that meteorology and emissions affect pNO3- by 35.3 % and 64.7 %, respectively, while atmospheric chemistry (48.0 %) and humidity (17.1 %) are the key drivers of its formation. Photochemistry promotes the formation of pNO3- in summer, whereas liquid-phase reactions dominate in winter at higher humidity levels (>60 %). The industry source (IS) (14.3 %), combustion source (CS) (12.8 %), and transportation source (TS) (11.8 %) are the main emission sources. The formation of pNO3- by the primary emissions and the transformation of NOx emitted from CS and TS is more sensitive to the changes of meteorological conditions, and controlling CS has the greater benefits to reduce pNO3-. The proposed framework could provide a reliable method for identifying drivers of pNO3- pollution at different haze events, supporting the formulation of control measures.

Elevated serum levels of GPX4, NDUFS4, PRDX5, and TXNRD2 as predictive biomarkers for castration resistance in prostate cancer patients: an exploratory study

BACKGROUND

Prostate cancer (PCa) is a heterogeneous disease affecting over 14% of the male population worldwide. Although patients often respond positively to initial treatments within the first 2-3 years, many eventually develop a more lethal form of the disease known as castration-resistant PCa (CRPC). At present, no biomarkers that predict the onset of CRPC are available. This study aims to provide insights into the diagnosis and prediction of CRPC emergence.

METHODS

Protein expression dynamics were analysed in drug (androgen receptor inhibitor)-tolerant persister (DTP) and drug withdrawal cells using proteomics to identify potential biomarkers. These biomarkers were subsequently validated using a mouse model, 180-paired carcinoma/benign tissues, and 482 serum samples. Five machine learning algorithms were employed to build clinical prediction models, wherein the SHapley Additive exPlanation (SHAP) framework was used to interpret the best-performing model. Moreover, three regression models were developed to determine the Time from initial PCa diagnosis to CRPC development (TPC) in patients.

RESULTS

We identified that the protein expression levels of GPX4, NDUFS4, PRDX5, and TXNRD2 were significantly upregulated in PCa patients, particularly in those with CRPC. Among the tested machine learning models, the random forest and extreme gradient boosting models performed best on tissue and serum cohorts, achieving AUCs of 0.958 and 0.988, respectively. In addition, a significant inverse correlation was observed between TPC and serum levels of these four biomarkers. This correlation was formulated in three regression models, which achieved the smallest mean absolute error of 1.903 on independent datasets for predicting CRPC emergence.

CONCLUSION

Our study provides new insights into the role of DTP cells in CRPC development. The quad protein panel identified in our study, along with the post hoc and intrinsically explainable prediction models, may serve as a convenient and real-time prognostic tool, addressing the current lack of clinical biomarkers for CRPC.

Strengthening Discovery and Application of Artificial Intelligence in Anesthesiology: A Report from the Anesthesia Research Council

Interest in the potential applications of artificial intelligence in medicine, anesthesiology, and the world at large has never been higher. The Anesthesia Research Council steering committee formed an anesthesiologist artificial intelligence expert workgroup charged with evaluating the current state of artificial intelligence in anesthesiology, providing examples of future artificial intelligence applications and identifying barriers to artificial intelligence progress. The workgroup's findings are summarized here, starting with a brief introduction to artificial intelligence for clinicians, followed by overviews of current and anticipated artificial intelligence-focused research and applications in anesthesiology. Anesthesiology's progress in artificial intelligence is compared to that of other medical specialties, and barriers to artificial intelligence development and implementation in our specialty are discussed. The workgroup's recommendations address stakeholders in policymaking, research, development, implementation, training, and use of artificial intelligence-based tools for perioperative care.

Artificial intelligence in surgical medicine: a brief review

The application of artificial intelligence (AI) technology in the medical field, particularly in surgical operations, has evolved from science fiction to a crucial tool. With continuous advancements in computational power and algorithmic technology, AI is reshaping the surgical medicine landscape. From preoperative diagnosis and planning to intraoperative real-time navigation and assistance and postoperative rehabilitation and follow-up management, AI technology has significantly enhanced the precision and safety of surgical procedures. This paper systematically reviews the development and current applications of AI in surgery, focusing on specific case studies of AI in surgical procedures, diagnostic assistance, intraoperative navigation, and postoperative management, highlighting its significant contributions to improving surgical precision and safety. Despite the obvious advantages of AI in improving surgical success, reducing postoperative complications, and accelerating patient recovery, its use in surgery still faces numerous challenges, including its cost-effectiveness, dependency, data privacy and security, clinical integration, and physician training. This review summarizes the current applications of AI in surgical medicine, highlights its benefits and limitations, and discusses the challenges and future directions of integrating AI into surgical practice.

Current methods in explainable artificial intelligence and future prospects for integrative physiology

Explainable artificial intelligence (XAI) is gaining importance in physiological research, where artificial intelligence is now used as an analytical and predictive tool for many medical research questions. The primary goal of XAI is to make AI models understandable for human decision-makers. This can be achieved in particular through providing inherently interpretable AI methods or by making opaque models and their outputs transparent using post hoc explanations. This review introduces XAI core topics and provides a selective overview of current XAI methods in physiology. It further illustrates solved and discusses open challenges in XAI research using existing practical examples from the medical field. The article gives an outlook on two possible future prospects: (1) using XAI methods to provide trustworthy AI for integrative physiological research and (2) integrating physiological expertise about human explanation into XAI method development for useful and beneficial human-AI partnerships.

Associations between per- and polyfluoroalkyl chemicals and abdominal aortic calcification in middle-aged and older adults

INTRODUCTION

Per- and polyfluoroalkyl substances (PFAS) have infiltrated countless everyday products, raising concerns about potential effects on human health, specifically on the cardiovascular system and the development of abdominal aortic calcification (AAC). However, our understanding of this relationship is still limited.

OBJECTIVES

This study aims to investigate the effects of PFAS on AAC using machine learning algorithms.

METHODS

Leveraging the power of machine learning technique, extreme gradient boosting (XGBoost), we assessed the relationship between PFAS exposure and AAC risk. We focused on three PFAS compounds, perfluorodecanoic acid (PFDeA), perfluorohexane sulfonic acid (PFHxS), and perfluorononanoic acid (PFNA) through multiple logistic regression, restricted cubic spline (RCS), and quantile g-computation (QGC) models. To get more insight into the underlying mechanisms, mediation analyses are used to investigate the potential mediating role of fatty acids and blood cell fractions in AAC.

RESULTS

Our findings indicate that elevated serum levels of PFHxS and PFDeA are associated with the increased risk of AAC. The QGC analyses underscore the overall positive association between the PFAS mixture and AAC risk, with PFHxS carrying the greatest weight, followed by PFDeA. The RCS analyses reveal a dose-dependent increase between serum PFHxS concentration and AAC risk in an inverted V-shape way. Moreover, age and PFHxS exposure are identified as the primary factors contributing to abdominal aortic calcification risk in SHapley Additive exPlanation (SHAP) summary plot combined with XGBoost technique. Although PFAS significantly change the profile of fatty acids, we do not find any mediating roles of them in AAC. Despite strong associations between PFAS exposure and hematological indicators, our analysis does not find evidence that these indicators mediate the development of AAC.

CONCLUSIONS

In summary, our study highlights the detrimental impact of PFAS on abdominal aortic health and emphasizes the need for further research to understand the underlying mechanisms involved.

Strengthening Discovery and Application of Artificial Intelligence in Anesthesiology: A Report from the Anesthesia Research Council

Interest in the potential applications of artificial intelligence in medicine, anesthesiology, and the world at large has never been higher. The Anesthesia Research Council steering committee formed an anesthesiologist artificial intelligence expert workgroup charged with evaluating the current state of artificial intelligence in anesthesiology, providing examples of future artificial intelligence applications and identifying barriers to artificial intelligence progress. The workgroup's findings are summarized here, starting with a brief introduction to artificial intelligence for clinicians, followed by overviews of current and anticipated artificial intelligence-focused research and applications in anesthesiology. Anesthesiology's progress in artificial intelligence is compared to that of other medical specialties, and barriers to artificial intelligence development and implementation in our specialty are discussed. The workgroup's recommendations address stakeholders in policymaking, research, development, implementation, training, and use of artificial intelligence-based tools for perioperative care.

Development and Validation of a Novel Nomogram Risk Prediction Model for In-Hospital Death Following Extended Aortic Arch Repair for Acute Type A Aortic Dissection

Background

Extended aortic arch repair (EAR) is increasingly adopted for treating acute type A aortic dissection (ATAAD). However, existing prediction models may not be suitable for assessing the in-hospital death risk in ATAAD patients undergoing EAR. This study aims to develop a comprehensive risk prediction model for in-hospital death following EAR based on patient's preoperative status and surgical data, which may contribute to identification of high-risk individuals and improve outcomes following EAR.

Methods

We reviewed clinical records of consecutive adult ATAAD patients undergoing EAR at our institute between January 2015 and December 2022. Utilizing data from 925 ATAAD patients undergoing EAR, we employed multivariable logistic regression and machine learning techniques, respectively, to develop nomograms for in-hospital mortality. Employed machine learning techniques included simple decision tree, random forest (RF), eXtreme Gradient Boosting (XGBoost), and support vector machine (SVM).

Results

The nomogram based on SVM outperformed others, achieving a mean area under the receiver operating characteristic (ROC) curve (AUC) of 0.842 on training dataset and a mean AUC of 0.782 on testing dataset, accompanied by a Brier score of 0.058. Key risk factors included cerebral malperfusion, mesenteric malperfusion, preoperative critical station, Marfan syndrome, platelet count, D-dimer, coronary artery bypass grafting, and cardiopulmonary bypass time. A web-based application was developed for clinical use.

Conclusions

We develop a novel nomogram risk prediction model based on SVM algorithm for in-hospital death following extended aortic arch repair for ATAAD with good discrimination and accuracy.

Clinical Trial Registration

Registration number ChiCTR2200066414, https://www.chictr.org.cn/showproj.html?proj=187074.

The Inner Loop of Collective Human-Machine Intelligence

With the rise of artificial intelligence (AI) and the desire to ensure that such machines work well with humans, it is essential for AI systems to actively model their human teammates, a capability referred to as Machine Theory of Mind (MToM). In this paper, we introduce the inner loop of human-machine teaming expressed as communication with MToM capability. We present three different approaches to MToM: (1) constructing models of human inference with well-validated psychological theories and empirical measurements; (2) modeling human as a copy of the AI; and (3) incorporating well-documented domain knowledge about human behavior into the above two approaches. We offer a formal language for machine communication and MToM, where each term has a clear mechanistic interpretation. We exemplify the overarching formalism and the specific approaches in two concrete example scenarios. Related work that demonstrates these approaches is highlighted along the way. The formalism, examples, and empirical support provide a holistic picture of the inner loop of human-machine teaming as a foundational building block of collective human-machine intelligence.

Short-term peri- and intra-tumoral CT radiomics to predict immunotherapy response in advanced non-small cell lung cancer

Background

Predicting response to immunotherapy is crucial for advanced non-small cell lung cancer (NSCLC) treatment planning, but effective predictive markers for immunotherapy efficacy are still lacking. This study aimed to develop an explainable machine learning model for predicting immunotherapy responses in advanced NSCLC patients.

Methods

A total of 245 advanced NSCLC patients from two centers who received immunotherapy were retrospectively enrolled. For each primary tumor, three regions of interest were analyzed, namely, the intratumoral region (ITR), peritumoral region (PTR), and combined intratumoral and PTR (IPTR). Pre-radiomics features and delta-radiomics features reflecting the rate of change between radiomics features before and after treatment were extracted. Models for predicting immunotherapy responses were established via the extreme gradient boosting (XGBoost) classifier and assessed in terms of discrimination, calibration, and clinical utility. The SHapley Additive exPlanations (SHAP) tool was employed to explore the interpretability of the model. Kaplan-Meier (KM) analysis of progression-free survival (PFS) was conducted to evaluate the prognostic value of the prediction models.

Results

The delta-radiomics models of ITR and IPTR demonstrated optimal performance in predicting immunotherapy response, significantly improving the area under the curve (AUC) to 0.85 and 0.83 in the internal validation cohort and 0.84 and 0.86 in the external validation cohort. SHAP revealed a strong relationship between the delta-radiomics feature values and the model-predicted probabilities. KM curves indicated that the high-risk groups identified by the delta-radiomics models had significantly worse PFS than did the low-risk groups across all cohorts.

Conclusions

The results demonstrated that a model based on multiple time points outperformed one based on a single time point. The delta-radiomics model has been proved a noninvasive approach for assessing the response of advanced NSCLC patients to immunotherapy and facilitates individualized treatment decision making.

Integrated machine learning constructed a circadian-rhythm-related model to assess clinical outcomes and therapeutic advantages in hepatocellular carcinoma

Background

Circadian rhythm (CR) coordinates a variety of internal biological processes with the external daily cycles of light and dark. However, the implications of CR-related regulator in hepatocellular carcinoma (HCC) are quite obscure. Here, we aimed to identify pivotal CR-related markers in HCC for predicting survival and treatment outcomes.

Methods

The prognostic value of CR regulators in HCC was analyzed. Multi-step machine learning feature selection approaches were employed to establish a model. Thereafter, we evaluated its capacity of clinical prediction and treatment guidance.

Results

First, we depicted the prognostic stratification value of CR regulators in HCC. Two CR-related phenotypes were identified, revealing a distinct clinical outcome, biological pathways and drug sensitivity. Subsequently, via four topological approaches and differentially expressed genes (DEGs) from real-world cohorts, we screened out CRY2 as the pivotal CR regulator with significant prognostic value in HCC. We performed the relevant basic assay validation for CRY2. Overexpression of CRY2 inhibited the proliferation and migration abilities of Huh7 and Hep3B cells. Moreover, three machine learning algorithms [random forest (RF), extreme gradient boosting (XGBoost) and least absolute shrinkage and selection operator (LASSO)] were implemented to construct a risk-scoring model named CR predictor, which exhibited clinical benefits and therapeutic advantages for HCC. An online nomogram based on CR predictor was developed for predicting individualized survival (https://lihc.shinyapps.io/CR_predictor/). Finally, Mendelian randomization (MR) was performed. Among model genes in CR predictor, PPARGC1A revealed a significant causal effect on HCC.

Conclusions

We proposed a CR-related risk classifier in HCC, to predict patients' overall survival (OS) and therapeutic response. Targeting CR could be a promising treatment modality against HCC.

Predicting rheumatoid arthritis progression from seronegative undifferentiated arthritis using machine learning: a deep learning model trained on the KURAMA cohort and externally validated with the ANSWER cohort

BACKGROUND

Undifferentiated arthritis (UA) often develops into rheumatoid arthritis (RA), but predicting disease progression from seronegative UA remains challenging because seronegative RA often does not meet the classification criteria. This study aims to build a machine learning (ML) model to predict the progression from seronegative UA to RA using clinical and laboratory parameters.

METHODS

KURAMA cohort (training dataset) and ANSWER cohort (validation dataset) were utilized. Patients with seronegative UA were selected based on specific inclusion and exclusion criteria. Clinical and laboratory parameters, including demographic data, acute phase reactants, autoantibodies, and physical examination findings, were collected. Various ML models, including a Feedforward Neural Network (FNN), were developed and compared. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), sensitivity, and other metrics. SHapley Additive exPlanations (SHAP) values were computed to interpret the importance of variables.

RESULTS

KURAMA cohort included 210 patients with seronegative UA, of whom 57 (27.1%) progressed to RA. The FNN model demonstrated the highest predictive performance with an AUC of 0.924 and a sensitivity of 80.7% in the training dataset. Validation with ANSWER cohort (140 patients; 32.1% progressed to RA) showed an AUC of 0.777, sensitivity of 77.8%. MMP-3 had the highest impact on the model.

CONCLUSIONS

The FNN model exhibited robust performance in predicting the progression of RA from seronegative UA and maintained substantial sensitivity in an independent validation cohort. This model using only clinical and laboratory parameters has potential for predicting RA progression in patients with seronegative UA.

An interpreting machine learning models to predict amputation risk in patients with diabetic foot ulcers: a multi-center study

Background

Diabetic foot ulcers (DFUs) constitute a significant complication among individuals with diabetes and serve as a primary cause of nontraumatic lower-extremity amputation (LEA) within this population. We aimed to develop machine learning (ML) models to predict the risk of LEA in DFU patients and used SHapley additive explanations (SHAPs) to interpret the model.

Methods

In this retrospective study, data from 1,035 patients with DFUs at Sun Yat-sen Memorial Hospital were utilized as the training cohort to develop the ML models. Data from 297 patients across multiple tertiary centers were used for external validation. We then used least absolute shrinkage and selection operator analysis to identify predictors of amputation. We developed five ML models [logistic regression (LR), support vector machine (SVM), random forest (RF), k-nearest neighbors (KNN) and extreme gradient boosting (XGBoost)] to predict LEA in DFU patients. The performance of these models was evaluated using several metrics, including the area under the receiver operating characteristic curve (AUC), decision curve analysis (DCA), precision, recall, accuracy, and F1 score. Finally, the SHAP method was used to ascertain the significance of the features and to interpret the model.

Results

In the final cohort comprising 1332 individuals, 600 patients underwent amputation. Following hyperparameter optimization, the XGBoost model achieved the best amputation prediction performance with an accuracy of 0.94, a precision of 0.96, an F1 score of 0.94 and an AUC of 0.93 for the internal validation set on the basis of the 17 features. For the external validation set, the model attained an accuracy of 0.78, a precision of 0.93, an F1 score of 0.78, and an AUC of 0.83. Through SHAP analysis, we identified white blood cell counts, lymphocyte counts, and blood urea nitrogen levels as the model's main predictors.

Conclusion

The XGBoost algorithm-based prediction model can be used to dynamically estimate the risk of LEA in DFU patients, making it a valuable tool for preventing the progression of DFUs to amputation.

Exploring the role of breastfeeding, antibiotics, and indoor environments in preschool children atopic dermatitis through machine learning and hygiene hypothesis

The increasing global incidence of atopic dermatitis (AD) in children, especially in Western industrialized nations, has attracted considerable attention. The hygiene hypothesis, which posits that early pathogen exposure is crucial for immune system development, is central to understanding this trend. Furthermore, advanced machine learning algorithms have provided fresh insights into the interactions among various risk factors. This study investigates the relationship between early childhood antibiotic use, the duration of exclusive breastfeeding, indoor environmental factors, and child AD. By integrating machine learning techniques with the hygiene hypothesis, we aim to assess and interpret the significance of these risk factors. In this community-based case-control study with a 1:4 matching design, we evaluated the prevalence of AD in preschool-aged children. Data were collected via questionnaires completed by the parents of 771 children diagnosed with AD, matched with controls based on gender, age, and ethnicity. Univariate analyses identified relevant characteristics, which were further examined using multivariable logistic regression to calculate odds ratios (ORs). Stratified analyses assessed confounders and interactions, while the significance of variables was determined using a machine learning model. Renovating the dwelling during the mother's pregnancy (OR = 1.50; 95% CI 1.15-1.96) was identified as a risk factor for childhood AD. Additionally, antibiotic use three or more times during the child's first year (OR = 1.92; 95% CI 1.29-2.85) increased the risk of AD, independent of the parents' history of atopic disease and the child's mode of birth. Moreover, exclusive breastfeeding for four months or more (OR = 1.59; 95% CI 1.17-2.17) was identified as a risk factor for AD, particularly in the group without a maternal history of atopic disease. In contrast, having older siblings in the family (OR = 0.76; 95% CI 0.63-0.92) and low birth weight (OR = 0.62; 95% CI 0.47-0.81) were identified as protective factors against AD. Machine learning modeling indicated that the duration of exclusive breastfeeding, having older siblings, low birth weight, and parental history of AD or allergic rhinitis are key predictors of childhood AD. Our findings support the broader interpretation of the hygiene hypothesis. Machine learning analysis highlights the key role of the hygiene hypothesis and underscores the need for future AD prevention and healthcare initiatives focusing on children with a parental history of AD or allergic rhinitis. Moreover, minimizing antibiotic overuse may be essential for preventing AD in children. Further research is necessary to elucidate the impact and mechanisms of exclusive breastfeeding on AD to instruct maternal and child healthcare practices.

Comparison of 46 Cytokines in Peripheral Blood Between Patients with Papillary Thyroid Cancer and Healthy Individuals with AI-Driven Analysis to Distinguish Between the Two Groups

Background: Recent studies have analyzed some cytokines in patients with papillary thyroid carcinoma (PTC), but simultaneous analysis of multiple cytokines remains rare. Nonetheless, the simultaneous assessment of multiple cytokines is increasingly recognized as crucial for understanding the cytokine characteristics and developmental mechanisms in PTC. In addition, studies applying artificial intelligence (AI) to discriminate patients with PTC based on serum multiple cytokine data have been performed rarely. Here, we measured and compared 46 cytokines in patients with PTC and healthy individuals, applying AI algorithms to classify the two groups. Methods: Blood serum was isolated from 63 patients with PTC and 63 control individuals. Forty-six cytokines were analyzed simultaneously using Luminex assay Human XL Cytokine Panel. Several laboratory findings were identified from electronic medical records. Student's t-test or the Mann-Whitney U test were performed to analyze the difference between the two groups. As AI classification algorithms to categorize patients with PTC, K-nearest neighbor function, Naïve Bayes classifier, logistic regression, support vector machine, and eXtreme Gradient Boosting (XGBoost) were employed. The SHAP analysis assessed how individual parameters influence the classification of patients with PTC. Results: Cytokine levels, including GM-CSF, IFN-γ, IL-1ra, IL-7, IL-10, IL-12p40, IL-15, CCL20/MIP-α, CCL5/RANTES, and TNF-α, were significantly higher in PTC than in controls. Conversely, CD40 Ligand, EGF, IL-1β, PDGF-AA, and TGF-α exhibited significantly lower concentrations in PTC compared to controls. Among the five classification algorithms evaluated, XGBoost demonstrated superior performance in terms of accuracy, precision, sensitivity (recall), specificity, F1-score, and ROC-AUC score. Notably, EGF and IL-10 were identified as critical cytokines that significantly contributed to the differentiation of patients with PTC. Conclusions: A total of 5 cytokines showed lower levels in the PTC group than in the control, while 10 cytokines showed higher levels. While XGBoost demonstrated the best performance in discriminating between the PTC group and the control group, EGF and IL-10 were considered to be closely associated with PTC.

Machine Learning-Based Explainable Automated Nonlinear Computation Scoring System for Health Score and an Application for Prediction of Perioperative Stroke: Retrospective Study

BACKGROUND

Machine learning (ML) has the potential to enhance performance by capturing nonlinear interactions. However, ML-based models have some limitations in terms of interpretability.

OBJECTIVE

This study aimed to develop and validate a more comprehensible and efficient ML-based scoring system using SHapley Additive exPlanations (SHAP) values.

METHODS

We developed and validated the Explainable Automated nonlinear Computation scoring system for Health (EACH) framework score. We developed a CatBoost-based prediction model, identified key features, and automatically detected the top 5 steepest slope change points based on SHAP plots. Subsequently, we developed a scoring system (EACH) and normalized the score. Finally, the EACH score was used to predict perioperative stroke. We developed the EACH score using data from the Seoul National University Hospital cohort and validated it using data from the Boramae Medical Center, which was geographically and temporally different from the development set.

RESULTS

When applied for perioperative stroke prediction among 38,737 patients undergoing noncardiac surgery, the EACH score achieved an area under the curve (AUC) of 0.829 (95% CI 0.753-0.892). In the external validation, the EACH score demonstrated superior predictive performance with an AUC of 0.784 (95% CI 0.694-0.871) compared with a traditional score (AUC=0.528, 95% CI 0.457-0.619) and another ML-based scoring generator (AUC=0.564, 95% CI 0.516-0.612).

CONCLUSIONS

The EACH score is a more precise, explainable ML-based risk tool, proven effective in real-world data. The EACH score outperformed traditional scoring system and other prediction models based on different ML techniques in predicting perioperative stroke.

Using machine learning to predict depression among middle-aged and elderly population in China and conducting empirical analysis

OBJECTIVE

To develop a predictive model for evaluating depression among middle-aged and elderly individuals in China.

METHODS

Participants aged ≥ 45 from the 2020 China Health and Retirement Survey (CHARLS) cross-sectional study were enrolled. Depressive mood was defined as a score of 10 or higher on the CESD-10 scale, which has a maximum score of 30. A predictive model was developed using five selected machine learning algorithms. The model was trained and validated on the 2020 database cohort and externally validated through a questionnaire survey of middle-aged and elderly individuals in Shaanxi Province, China, following the same criteria. SHapley Additive Interpretation (SHAP) was employed to assess the importance of predictive factors.

RESULTS

The stacked ensemble model demonstrated an AUC of 0.8021 in the test set of the training cohort for predicting depressive symptoms; the corresponding AUC in the external validation cohort was 0.7448, outperforming all base models.

CONCLUSION

The stacked ensemble approach serves as an effective tool for identifying depression in a large population of middle-aged and elderly individuals in China. For depression prediction, factors such as life satisfaction, self-reported health, pain, sleep duration, and cognitive function are identified as highly significant predictive factors.

Optimal Vasopressin Initiation in Septic Shock: The OVISS Reinforcement Learning Study

Importance

Norepinephrine is the first-line vasopressor for patients with septic shock. When and whether a second agent, such as vasopressin, should be added is unknown.

Objective

To derive and validate a reinforcement learning model to determine the optimal initiation rule for vasopressin in adult, critically ill patients receiving norepinephrine for septic shock.

Design, Setting, and Participants

Reinforcement learning was used to generate the optimal rule for vasopressin initiation to improve short-term and hospital outcomes, using electronic health record data from 3608 patients who met the Sepsis-3 shock criteria at 5 California hospitals from 2012 to 2023. The rule was evaluated in 628 patients from the California dataset and 3 external datasets comprising 10 217 patients from 227 US hospitals, using weighted importance sampling and pooled logistic regression with inverse probability weighting.

Exposures

Clinical, laboratory, and treatment variables grouped hourly for 120 hours in the electronic health record.

Main Outcome and Measure

The primary outcome was in-hospital mortality.

Results

The derivation cohort (n = 3608) included 2075 men (57%) and had a median (IQR) age of 63 (56-70) years and Sequential Organ Failure Assessment (SOFA) score at shock onset of 5 (3-7 [range, 0-24, with higher scores associated with greater mortality]). The validation cohorts (n = 10 217) were 56% male (n = 5743) with a median (IQR) age of 67 (57-75) years and a SOFA score of 6 (4-9). In validation data, the model suggested vasopressin initiation in more patients (87% vs 31%), earlier relative to shock onset (median [IQR], 4 [1-8] vs 5 [1-14] hours), and at lower norepinephrine doses (median [IQR], 0.20 [0.08-0.45] vs 0.37 [0.17-0.69] µg/kg/min) compared with clinicians' actions. The rule was associated with a larger expected reward in validation data compared with clinician actions (weighted importance sampling difference, 31 [95% CI, 15-52]). The adjusted odds of hospital mortality were lower if vasopressin initiation was similar to the rule compared with different (odds ratio, 0.81 [95% CI, 0.73-0.91]), a finding consistent across external validation sets.

Conclusions and Relevance

In adult patients with septic shock receiving norepinephrine, the use of vasopressin was variable. A reinforcement learning model developed and validated in several observational datasets recommended more frequent and earlier use of vasopressin than average care patterns and was associated with reduced mortality.

A clinical data-driven machine learning approach for predicting the effectiveness of piperacillin-tazobactam in treating lower respiratory tract infections

BACKGROUND

In hospitalized patients, inadequate antibiotic dosage leading to bacterial resistance and increased antimicrobial use intensity due to overexposure to antibiotics are common problems. In the present study, we constructed a machine learning model based on patients' clinical information to predict the clinical effectiveness of Piperacillin-tazobactam (TZP) (4:1) in treating bacterial lower respiratory tract infections (LRTIs), to assist clinicians in making better clinical decisions.

METHODS

We collected data from patients diagnosed with LRTIs or equivalent diagnoses admitted to the Department of Pulmonary and Critical Care Medicine at Shanghai Pudong Hospital, Shanghai, between January 1, 2021, and July 31, 2023. A total of 26 relevant clinical features were extracted from this cohort. Following data preprocessing, we trained four models: Logistic Regression, Random Forest, Support Vector Machine, and Gaussian Naive Bayes. The dataset was split into training and test sets using a 7:3 ratio. The top-performing models, as determined by Receiver Operating Characteristic (ROC)-Area Under the Curve (AUC) on the independent test set, were subsequently ensembled. Ensemble model (EL) performance was evaluated using bootstrap resampling on the training set and ROC-AUC, recall, accuracy, precision, F1-score, and log loss on an independent test set. The optimal model was then deployed as a web application for clinical outcome prediction.

RESULTS

A total of 1,314 patients primarily treated with TZP as initial empiric antibiotic therapy were enrolled in the analysis. The success group comprised 995 patients (75.7%), while the failure group consisted of 319 patients (24.3%). We constructed an ensemble learning model based on the Logistic Regression, Support Vector Machine and Random Forest models, which showed better overall performance. The EL model demonstrated robust performance on an independent test set, exhibiting a ROC-AUC of 0.69, a recall of 0.69, an accuracy of 0.64, a precision of 0.40, a F1-score of 0.50, and a log loss of 0.66. A corresponding web application was then developed and made available at http://106.12.146.54:1020/ .

CONCLUSIONS

In this study, we successfully developed and validated an EL model that effectively predicts the clinical effectiveness of TZP (4:1) in treating bacterial LRTIs. The model achieved a balanced performance across key evaluation metrics, demonstrating the model's potential utility in clinical decision-making. The web-based application makes this model readily accessible to clinicians, potentially helping optimize antibiotic dosing decisions and reduce both inadequate treatment and overexposure. While promising, future studies with larger datasets and prospective validation are needed to further improve the model's performance and validate its clinical utility. This work represents a step forward in using machine learning to support antimicrobial stewardship and personalized antibiotic therapy.