An artificial neural network-pharmacokinetic model and its interpretation using Shapley additive explanations

Value of Bioinformatics Models for Predicting Translational Control of Angiogenesis

Angiogenesis, the formation of new blood vessels, is a fundamental biological process with implications for both physiological functions and pathological conditions. While the transcriptional regulation of angiogenesis, mediated by factors such as HIF-1α (hypoxia-inducible factor 1-alpha) and VEGF (vascular endothelial growth factor), is well-characterized, the translational regulation of this process remains underexplored. Bioinformatics has emerged as an indispensable tool for advancing our understanding of translational regulation, offering predictive models that leverage large data sets to guide research and optimize experimental approaches. However, a significant gap persists between bioinformatics experts and other researchers, limiting the accessibility and utility of these tools in the broader scientific community. To address this divide, user-friendly bioinformatics platforms are being developed to democratize access to predictive analytics and empower researchers across disciplines. Translational control, compared with transcriptional control, offers a more energy-efficient mechanism that facilitates rapid cellular responses to environmental changes. Furthermore, transcriptional regulators themselves are often subject to translational control, emphasizing the interconnected nature of these regulatory layers. Investigating translational regulation requires advanced, accessible bioinformatics tools to analyze RNA structures, interacting micro-RNAs, long noncoding RNAs, and RBPs (RNA-binding proteins). Predictive platforms such as RNA structure, human internal ribosome entry site Atlas, and RBPSuite enable the study of RNA motifs and RNA-protein interactions, shedding light on these critical regulatory mechanisms. This review highlights the transformative role of bioinformatics using widely accessible user-friendly tools with a Web-browser interface to elucidate translational regulation in angiogenesis. The bioinformatics tools discussed extend beyond angiogenesis, with applications in diverse fields, including clinical care. By integrating predictive models and experimental insights, researchers can streamline hypothesis generation, reduce experimental costs, and find novel translational regulators. By bridging the bioinformatics knowledge gap, this review aims to empower researchers worldwide to adopt bioinformatics tools in their work, fostering innovation and accelerating scientific discovery.

Machine Learning for Prediction of Drug Concentrations: Application and Challenges

With the advancements in algorithms and increased accessibility of multi-source data, machine learning in pharmacokinetics is gaining interest. This review summarizes studies on machine learning-based pharmacokinetics analysis up to September 2024, identified from the PubMed and IEEE Xplore databases. The main focus of this review is on the use of machine learning in predicting drug concentration. This review provides a comprehensive summary of the advances in the machine learning algorithms for pharmacokinetics analysis. Specifically, we describe the common practices in data preprocessing, the application scenarios of various algorithms, and the critical challenges that require attention. Most machine learning models show comparable performance to those of population pharmacokinetics models. Tree-based algorithms and neural networks have the most applications. Furthermore, the use of ensemble modeling techniques can improve the accuracy of these models' predictions of drug concentrations, especially the ensembles of machine learning and pharmacometrics.

Machine learning-based prediction of contralateral knee osteoarthritis development using the Osteoarthritis Initiative and the Multicenter Osteoarthritis Study dataset

Having osteoarthritis in one knee is reported as an independent risk factor for developing contralateral knee osteoarthritis (KOA). However, no study has been designed to predict the development of contralateral KOA (cKOA). The authors hypothesized that specific risk factors for cKOA development exist and that it could be accurately predicted with the assistance of machine learning. KOA was defined using the Kellgren-Lawrence grade (KLG) of 2 or higher. Data from 1353 unilateral KOA patients (900 from the Osteoarthritis Initiative [OAI] and 453 from the Multicenter Osteoarthritis Study [MOST]) over 4-5 years of follow-up were examined. The risk factors for cKOA development were analyzed, and a machine learning model was developed to predict cKOA using OAI as the development data set and MOST as the test data set. cKOA developed in 172 (19.1%) and 178 (39.3%) of the patients (OAI and MOST, respectively) over a period of 4-5 years. A machine learning model was developed using the Tree-based Pipeline Optimization Tool algorithm. This model utilized nine variables, including baseline lateral joint space narrowing grade of the contralateral knee (odds ratio 4.475). The area under the curve of the receiver operating characteristics curve, along with accuracy, precision, and F1-score, were recorded as 0.69, 0.60, 0.50, and 0.58, respectively, in the test data set. The development of cKOA could be effectively predicted using a limited number of variables through machine learning. Surgeons should consider the development of cKOA in patients with identified risk factors when managing KOA patients.

Comparison of population pharmacokinetic modeling and machine learning approaches for predicting voriconazole trough concentrations in critically ill patients

BACKGROUND

Despite the widespread use of voriconazole in antifungal treatment, its high pharmacokinetic and pharmacodynamic variability may lead to suboptimal efficacy, especially in intensive care unit (ICU) patients. Machine learning (ML), an artificial intelligence modeling approach, is increasingly being applied to personalized medicine. The effectiveness of ML models for predicting voriconazole blood concentrations in ICU patients, compared to traditional population pharmacokinetics (popPK) models, has been uncertain until now. This study aims to identify the most effective modeling strategy for voriconazole.

METHODS

We developed six ML models using 244 concentrations from 62 patients in our previous popPK dataset. Another additional dataset, consisting of 282 trough concentrations from 177 patients, was used to externally evaluate both ML models and five other published popPK models, utilizing prediction-based diagnostics, simulation-based diagnostics, and Bayesian forecasting.

RESULTS

The XGBoost model exhibited superior predictive performance among the six ML models, achieving an R2 of 0.73. Its performance metrics (RMSE%: 127.21 %, median absolute prediction error: 29.65 %, median prediction error: 9.82 %, F20: 34.04 %, F30: 50.71 %) outperformed those of the best popPK model (RMSE%: 152.41 %, median absolute prediction error: 44.75 %, median prediction error: -0.99 %, F20: 23.40 %, F30: 36.88 %), suggesting greater accuracy and precision in predicting pharmacokinetics.

CONCLUSIONS

Both ML and popPK models can be utilized for individualized voriconazole therapy. Our comparative study provides insights into the most effective methods for modeling and predicting voriconazole concentrations.

Influence of Processing Parameters on Mechanical Properties and Degree of Crystallization of Polylactide

This work attempts to assess the influence of process parameters on the change of mechanical properties and the degree of crystallinity of polylactide (PLA). PLA is a biodegradable material that has been widely used in various areas-from packaging, through medicine, to 3D printing, where it is used to produce prototypes. The method of processing is important, because the technological process and its parameters have a significant impact on the quality of the finished product. Their appropriate selection depends on quality and mechanical properties. The process parameters have an impact on the structure of PLA, specifically on the share of the crystalline phase, which is also important from the point of view of the functional properties of the finished product. This work assessed the impact of the technological parameters of the injection process on the final properties of the obtained samples. The obtained results of static tensile strength, hardness and differential scanning calorimetry (DSC) analysis confirm that changing these parameters affects the material properties.

Implementing Explainable Machine Learning Models for Practical Prediction of Early Neonatal Hypoglycemia

Hypoglycemia is a common metabolic disorder that occurs in the neonatal period. Early identification of neonates at risk of developing hypoglycemia can optimize therapeutic strategies in neonatal care. This study aims to develop a machine learning model and implement a predictive application to assist clinicians in accurately predicting the risk of neonatal hypoglycemia within four hours after birth. Our retrospective study analyzed data from neonates born ≥35 weeks gestational age and admitted to the well-baby nursery between 1 January 2011 and 31 August 2021. We collected electronic medical records of 2687 neonates from a tertiary medical center in Southern Taiwan. Using 12 clinically relevant features, we evaluated nine machine learning approaches to build the predictive models. We selected the models with the highest area under the receiver operating characteristic curve (AUC) for integration into our hospital information system (HIS). The top three AUC values for the early neonatal hypoglycemia prediction models were 0.739 for Stacking, 0.732 for Random Forest and 0.732 for Voting. Random Forest is considered the best model because it has a relatively high AUC and shows no significant overfitting (accuracy of 0.658, sensitivity of 0.682, specificity of 0.649, F1 score of 0.517 and precision of 0.417). The best model was incorporated in the web-based application integrated into the hospital information system. Shapley Additive Explanation (SHAP) values indicated mode of delivery, gestational age, multiparity, respiratory distress, and birth weight < 2500 gm as the top five predictors of neonatal hypoglycemia. The implementation of our machine learning model provides an effective tool that assists clinicians in accurately identifying at-risk neonates for early neonatal hypoglycemia, thereby allowing timely interventions and treatments.

Performance Characterization and Composition Design Using Machine Learning and Optimal Technology for Slag-Desulfurization Gypsum-Based Alkali-Activated Materials

Fly ash-slag-based alkali-activated materials have excellent mechanical performance and a low carbon footprint, and they have emerged as a promising alternative to Portland cement. Therefore, replacing traditional Portland cement with slag-desulfurization gypsum-based alkali-activated materials will help to make better use of the waste, protect the environment, and improve the materials' performance. In order to better understand it and thus better use it in engineering, it needs to be characterized for performance and compositional design. This study developed a novel framework for performance characterization and composition design by combining Categorical Gradient Boosting (CatBoost), simplicial homology global optimization (SHGO), and laboratory tests. The CatBoost characterization model was evaluated and discussed based on SHapley Additive exPlanations (SHAPs) and a partial dependence plot (PDP). Through the proposed framework, the optimal composition of the slag-desulfurization gypsum-based alkali-activated materials with the maximum flexural strength and compressive strength at 1, 3, and 7 days is Ca(OH)2: 3.1%, fly ash: 2.6%, DG: 0.53%, alkali: 4.3%, modulus: 1.18, and W/G: 0.49. Compared with the material composition obtained from the traditional experiment, the actual flexural strength and compressive strength at 1, 3, and 7 days increased by 26.67%, 6.45%, 9.64%, 41.89%, 9.77%, and 7.18%, respectively. In addition, the results of the optimal composition obtained by laboratory tests are very close to the predictions of the developed framework, which shows that CatBoost characterizes the performance well based on test data. The developed framework provides a reasonable, scientific, and helpful way to characterize the performance and determine the optimal composition for civil materials.

Optimal use of β-lactams in neonates: machine learning-based clinical decision support system

BACKGROUND

Accurate prediction of the optimal dose for β-lactam antibiotics in neonatal sepsis is challenging. We aimed to evaluate whether a reliable clinical decision support system (CDSS) based on machine learning (ML) can assist clinicians in making optimal dose selections.

METHODS

Five β-lactam antibiotics (amoxicillin, ceftazidime, cefotaxime, meropenem and latamoxef), commonly used to treat neonatal sepsis, were selected. The CDSS was constructed by incorporating the drug, patient, dosage, pharmacodynamic, and microbiological factors. The CatBoost ML algorithm was used to build the CDSS. Real-world studies were used to evaluate the CDSS performance. Virtual trials were used to compare the CDSS-optimized doses with guideline-recommended doses.

FINDINGS

For a specific drug, by entering the patient characteristics and pharmacodynamic (PD) target (50%/70%/100% fraction of time that the free drug concentration is above the minimal inhibitory concentration [fT > MIC]), the CDSS can determine whether the planned dosing regimen will achieve the PD target and suggest an optimal dose. The prediction accuracy of all five drugs was >80.0% in the real-world validation. Compared with the PopPK model, the overall accuracy, precision, recall, and F1-Score improved by 10.7%, 22.1%, 64.2%, and 43.1%, respectively. Using the CDSS-optimized doses, the average probability of target concentration attainment increased by 58.2% compared to the guideline-recommended doses.

INTERPRETATION

An ML-based CDSS was successfully constructed to assist clinicians in selecting optimal β-lactam antibiotic doses.

FUNDING

This work was supported by the National Natural Science Foundation of China; Distinguished Young and Middle-aged Scholar of Shandong University; National Key Research and Development Program of China.

Machine Learning Approach in Dosage Individualization of Isoniazid for Tuberculosis

INTRODUCTION

Isoniazid is a first-line antituberculosis agent with high variability, which would profit from individualized dosing. Concentrations of isoniazid at 2 h (C2h), as an indicator of safety and efficacy, are important for optimizing therapy.

OBJECTIVE

The objective of this study was to establish machine learning (ML) models to predict the C2h, that can be used for establishing an individualized dosing regimen in clinical practice.

METHODS

Published population pharmacokinetic (PopPK) models for adults were searched based on PubMed and ultimately four reliable models were selected for simulating individual C2h datasets under different conditions (demographics, genotype, ethnicity, etc.). Machine learning models were trained on simulated C2h obtained from the four PopPK models. Five different algorithms were used for ML model building to predict C2h. Real-world data were used for predictive performance evaluations. Virtual trials were used to compare ML-optimized doses with PopPK model-optimized doses.

RESULTS

Categorical boosting (CatBoost) exhibited the highest prediction ability. Target C2h can be predicted using the ML model combined with the dosing regimen and three covariates (N-acetyltransferase 2 [NAT2] genotypes, weight and race [Asians and Africans]). Real-world data validation results showed that the ML model can achieve an overall prediction accuracy of 93.4%. Using the final ML model, the mean absolute prediction error value decreased by 45.7% relative to the average of PopPK models. Using the ML-optimized dosing regimen, the probability of target attainment increased by 43.7% relative to the PopPK model-optimized dosing regimens.

CONCLUSION

Machine learning models were developed with great predictive performance, which can be used to determine the individualized initial dose of isoniazid in adult patients.

Application of machine learning techniques in population pharmacokinetics/pharmacodynamics modeling

Population pharmacokinetics/pharmacodynamics (pop-PK/PD) consolidates pharmacokinetic and pharmacodynamic data from many subjects to understand inter- and intra-individual variability due to patient backgrounds, including disease state and genetics. The typical workflow in pop-PK/PD analysis involves the determination of the structure model, selection of the error model, analysis based on the base model, covariate modeling, and validation of the final model. Machine learning is gaining considerable attention in the medical and various fields because, in contrast to traditional modeling, which often assumes linear or predefined relationships, machine learning modeling learns directly from data and accommodates complex patterns. Machine learning has demonstrated excellent capabilities for prescreening covariates and developing predictive models. This review introduces various applications of machine learning techniques in pop-PK/PD research.

Warfarin-A natural anticoagulant: A review of research trends for precision medication

BACKGROUND

Warfarin is a widely prescribed anticoagulant in the clinic. It has a more considerable individual variability, and many factors affect its variability. Mathematical models can quantify the quantitative impact of these factors on individual variability.

PURPOSE

The aim is to comprehensively analyze the advanced warfarin dosing algorithm based on pharmacometrics and machine learning models of personalized warfarin dosage.

METHODS

A bibliometric analysis of the literature retrieved from PubMed and Scopus was performed using VOSviewer. The relevant literature that reported the precise dosage of warfarin calculation was retrieved from the database. The multiple linear regression (MLR) algorithm was excluded because a recent systematic review that mainly reviewed this algorithm has been reported. The following terms of quantitative systems pharmacology, mechanistic model, physiologically based pharmacokinetic model, artificial intelligence, machine learning, pharmacokinetic, pharmacodynamic, pharmacokinetics, pharmacodynamics, and warfarin were added as MeSH Terms or appearing in Title/Abstract into query box of PubMed, then humans and English as filter were added to retrieve the literature.

RESULTS

Bibliometric analysis revealed important co-occuring MeShH and index keywords. Further, the United States, China, and the United Kingdom were among the top countries contributing in this domain. Some studies have established personalized warfarin dosage models using pharmacometrics and machine learning-based algorithms. There were 54 related studies, including 14 pharmacometric models, 31 artificial intelligence models, and 9 model evaluations. Each model has its advantages and disadvantages. The pharmacometric model contains biological or pharmacological mechanisms in structure. The process of pharmacometric model development is very time- and labor-intensive. Machine learning is a purely data-driven approach; its parameters are more mathematical and have less biological interpretation. However, it is faster, more efficient, and less time-consuming. Most published models of machine learning algorithms were established based on cross-sectional data sourced from the database.

CONCLUSION

Future research on personalized warfarin medication should focus on combining the advantages of machine learning and pharmacometrics algorithms to establish a more robust warfarin dosage algorithm. Randomized controlled trials should be performed to evaluate the established algorithm of warfarin dosage. Moreover, a more user-friendly and accessible warfarin precision medicine platform should be developed.

Artificial Neural Network Analysis of Determinants of Tacrolimus Pharmacokinetics in Liver Transplant Recipients

BACKGROUND

The efficacy and toxicity of tacrolimus are closely related to its trough blood concentrations. Identifying the influencing factors of pharmacokinetics of tacrolimus in the early postoperative period is conducive to the optimization of the individualized tacrolimus administration protocol and to help liver transplant (LT) recipients achieve the target blood concentrations.

OBJECTIVE

This study aimed to develop an artificial neural network (ANN) for predicting the blood concentration of tacrolimus soon after liver transplantation and for identifying determinants of the concentration based on Shapley additive explanation (SHAP).

METHODS

In this retrospective study, we enrolled 31 recipients who were first treated with liver transplantation from the Department of Liver Transplantation and Hepatic Surgery, the First Affiliated Hospital of Shandong First Medical University (Shandong Provincial Qianfoshan Hospital) from November 2020 to May 2021. The basic information, biochemical indexes, use of concomitant drugs, and genetic factors of organ donors and recipients were used for the ANN model inputs, and the output was the steady-state trough concentration (C0) of tacrolimus after oral administration in LT recipients. The ANN model was established to predict C0 of tacrolimus, SHAP was applied to the trained model, and the SHAP value of each input was calculated to analyze quantitatively the influencing factors for the output C0.

RESULTS

A back-propagation ANN model with 3 hidden layers was established using deep learning. The mean prediction error was 0.27 ± 0.75 ng/mL; mean absolute error, 0.60 ± 0.52 ng/mL; correlation coefficient between predicted and actual C0 values, 0.9677; and absolute prediction error of all blood concentrations obtained by the ANN model, ≤3.0 ng/mL. The results indicated that the following factors had the most significant effect on C0: age, daily drug dose, genotype at CYP3A5 polymorphism rs776746 in both recipient and donor, and concomitant use of caspofungin. The predicted C0 value of tacrolimus in LT recipients increased in a dose-dependent manner when the daily dose exceeded 3 mg, whereas it decreased with age when LT recipients were older than 48 years. The predicted C0 was higher when recipients and donors had the genotype CYP3A5*3*3 than when they had the genotype CYP3A5*1. The predicted C0 value also increased with the use of caspofungin or Wuzhi capsule.

CONCLUSION AND RELEVANCE

The established ANN model can be used to predict the C0 value of tacrolimus in LT recipients with high accuracy and good predictive ability, serving as a reference for personalized treatment in the early stage after liver transplantation.

Machine Learning: A New Approach for Dose Individualization

The application of machine learning (ML) has shown promising results in precision medicine due to its exceptional performance in dealing with complex multidimensional data. However, using ML for individualized dosing of medicines is still in its early stage, meriting further exploration. A systematic review of study designs and modeling details of using ML for individualized dosing of different drugs was performed. We have summarized the status of the study populations, predictive targets, and data sources for ML modeling, the selection of ML algorithms and features, and the evaluation and validation of their predictive performance. We also used the Prediction model Risk of Bias Assessment Tool (PROBAST) to assess the risk of bias of included studies. Currently, ML can be used for both a priori and a posteriori dose selection and optimization, and it can also assist the implementation of therapeutic drug monitoring. However, studies are mainly focused on drugs with narrow therapeutic windows, predominantly immunosuppressants (N = 23, 35.9%) and anti-infectives (N = 21, 32.8%), and there is currently only very limited attention for special populations, such as children (N = 22, 34.4%). Most studies showed poor methodological quality and a high risk of bias. The lack of external validation and clinical utility evaluation currently limits the further clinical implementation of ML for dose individualization. We therefore have proposed several ways to improve the clinical relevance of the studies and facilitate the translation of ML models into clinical practice.

Interpretable machine learning model to predict surgical difficulty in laparoscopic resection for rectal cancer

Background

Laparoscopic total mesorectal excision (LaTME) is standard surgical methods for rectal cancer, and LaTME operation is a challenging procedure. This study is intended to use machine learning to develop and validate prediction models for surgical difficulty of LaTME in patients with rectal cancer and compare these models' performance.

Methods

We retrospectively collected the preoperative clinical and MRI pelvimetry parameter of rectal cancer patients who underwent laparoscopic total mesorectal resection from 2017 to 2022. The difficulty of LaTME was defined according to the scoring criteria reported by Escal. Patients were randomly divided into training group (80%) and test group (20%). We selected independent influencing features using the least absolute shrinkage and selection operator (LASSO) and multivariate logistic regression method. Adopt synthetic minority oversampling technique (SMOTE) to alleviate the class imbalance problem. Six machine learning model were developed: light gradient boosting machine (LGBM); categorical boosting (CatBoost); extreme gradient boost (XGBoost), logistic regression (LR); random forests (RF); multilayer perceptron (MLP). The area under receiver operating characteristic curve (AUROC), accuracy, sensitivity, specificity and F1 score were used to evaluate the performance of the model. The Shapley Additive Explanations (SHAP) analysis provided interpretation for the best machine learning model. Further decision curve analysis (DCA) was used to evaluate the clinical manifestations of the model.

Results

A total of 626 patients were included. LASSO regression analysis shows that tumor height, prognostic nutrition index (PNI), pelvic inlet, pelvic outlet, sacrococcygeal distance, mesorectal fat area and angle 5 (the angle between the apex of the sacral angle and the lower edge of the pubic bone) are the predictor variables of the machine learning model. In addition, the correlation heatmap shows that there is no significant correlation between these seven variables. When predicting the difficulty of LaTME surgery, the XGBoost model performed best among the six machine learning models (AUROC=0.855). Based on the decision curve analysis (DCA) results, the XGBoost model is also superior, and feature importance analysis shows that tumor height is the most important variable among the seven factors.

Conclusions

This study developed an XGBoost model to predict the difficulty of LaTME surgery. This model can help clinicians quickly and accurately predict the difficulty of surgery and adopt individualized surgical methods.

Training the next generation of pharmacometric modelers: a multisector perspective

The current demand for pharmacometricians outmatches the supply provided by academic institutions and considerable investments are made to develop the competencies of these scientists on-the-job. Even with the observed increase in academic programs related to pharmacometrics, this need is unlikely to change in the foreseeable future, as the demand and scope of pharmacometrics applications keep expanding. Further, the field of pharmacometrics is changing. The field largely started when Lewis Sheiner and Stuart Beal published their seminal papers on population pharmacokinetics in the late 1970's and early 1980's and has continued to grow in impact and use since its inception. Physiological-based pharmacokinetics and systems pharmacology have grown rapidly in scope and impact in the last decade and machine learning is just on the horizon. While all these methodologies are categorized as pharmacometrics, no one person can be an expert in everything. So how do you train future pharmacometricians? Leading experts in academia, industry, contract research organizations, clinical medicine, and regulatory gave their opinions on how to best train future pharmacometricians. Their opinions were collected and synthesized to create some general recommendations.

Development of a prediction model for the risk of 30-day unplanned readmission in older patients with heart failure: A multicenter retrospective study

BACKGROUND AND AIM

Heart failure (HF) imposes significant global health costs due to its high incidence, readmission, and mortality rate. Accurate assessment of readmission risk and precise interventions have become important measures to improve health for patients with HF. Therefore, this study aimed to develop a machine learning (ML) model to predict 30-day unplanned readmissions in older patients with HF.

METHODS AND RESULTS

This study collected data on hospitalized older patients with HF from the medical data platform of Chongqing Medical University from January 1, 2012, to December 31, 2021. A total of 5 candidate algorithms were selected from 15 ML algorithms with excellent performance, which was evaluated by area under the operating characteristic curve (AUC) and accuracy. Then, the 5 candidate algorithms were hyperparameter tuned by 5-fold cross-validation grid search, and performance was evaluated by AUC, accuracy, sensitivity, specificity, and recall. Finally, an optimal ML model was constructed, and the predictive results were explained using the SHapley Additive exPlanations (SHAP) framework. A total of 14,843 older patients with HF were consecutively enrolled. CatBoost model was selected as the best prediction model, and AUC was 0.732, with 0.712 accuracy, 0.619 sensitivity, and 0.722 specificity. NT.proBNP, length of stay (LOS), triglycerides, blood phosphorus, blood potassium, and lactate dehydrogenase had the greatest effect on 30-day unplanned readmission in older patients with HF, according to SHAP results.

CONCLUSIONS

The study developed a CatBoost model to predict the risk of unplanned 30-day special-cause readmission in older patients with HF, which showed more significant performance compared with the traditional logistic regression model.

An artificial intelligence-assisted physiologically-based pharmacokinetic model to predict nanoparticle delivery to tumors in mice

The critical barrier for clinical translation of cancer nanomedicine stems from the inefficient delivery of nanoparticles (NPs) to target solid tumors. Rapid growth of computational power, new machine learning and artificial intelligence (AI) approaches provide new tools to address this challenge. In this study, we established an AI-assisted physiologically based pharmacokinetic (PBPK) model by integrating an AI-based quantitative structure-activity relationship (QSAR) model with a PBPK model to simulate tumor-targeted delivery efficiency (DE) and biodistribution of various NPs. The AI-based QSAR model was developed using machine learning and deep neural network algorithms that were trained with datasets from a published "Nano-Tumor Database" to predict critical input parameters of the PBPK model. The PBPK model with optimized NP cellular uptake kinetic parameters was used to predict the maximum delivery efficiency (DEmax) and DE at 24 (DE24) and 168 h (DE168) of different NPs in the tumor after intravenous injection and achieved a determination coefficient of R2 = 0.83 [root mean squared error (RMSE) = 3.01] for DE24, R2 = 0.56 (RMSE = 2.27) for DE168, and R2 = 0.82 (RMSE = 3.51) for DEmax. The AI-PBPK model predictions correlated well with available experimentally-measured pharmacokinetic profiles of different NPs in tumors after intravenous injection (R2 ≥ 0.70 for 133 out of 288 datasets). This AI-based PBPK model provides an efficient screening tool to rapidly predict delivery efficiency of a NP based on its physicochemical properties without relying on an animal training dataset.

Use of Machine Learning for Dosage Individualization of Vancomycin in Neonates

BACKGROUND AND OBJECTIVE

High variability in vancomycin exposure in neonates requires advanced individualized dosing regimens. Achieving steady-state trough concentration (C0) and steady-state area-under-curve (AUC0-24) targets is important to optimize treatment. The objective was to evaluate whether machine learning (ML) can be used to predict these treatment targets to calculate optimal individual dosing regimens under intermittent administration conditions.

METHODS

C0 were retrieved from a large neonatal vancomycin dataset. Individual estimates of AUC0-24 were obtained from Bayesian post hoc estimation. Various ML algorithms were used for model building to C0 and AUC0-24. An external dataset was used for predictive performance evaluation.

RESULTS

Before starting treatment, C0 can be predicted a priori using the Catboost-based C0-ML model combined with dosing regimen and nine covariates. External validation results showed a 42.5% improvement in prediction accuracy by using the ML model compared with the population pharmacokinetic model. The virtual trial showed that using the ML optimized dose; 80.3% of the virtual neonates achieved the pharmacodynamic target (C0 in the range of 10-20 mg/L), much higher than the international standard dose (37.7-61.5%). Once therapeutic drug monitoring (TDM) measurements (C0) in patients have been obtained, AUC0-24 can be further predicted using the Catboost-based AUC-ML model combined with C0 and nine covariates. External validation results showed that the AUC-ML model can achieve an prediction accuracy of 80.3%.

CONCLUSION

C0-based and AUC0-24-based ML models were developed accurately and precisely. These can be used for individual dose recommendations of vancomycin in neonates before treatment and dose revision after the first TDM result is obtained, respectively.

OEDL: an optimized ensemble deep learning method for the prediction of acute ischemic stroke prognoses using union features

Background

Early stroke prognosis assessments are critical for decision-making regarding therapeutic intervention. We introduced the concepts of data combination, method integration, and algorithm parallelization, aiming to build an integrated deep learning model based on a combination of clinical and radiomics features and analyze its application value in prognosis prediction.

Methods

The research steps in this study include data source and feature extraction, data processing and feature fusion, model building and optimization, model training, and so on. Using data from 441 stroke patients, clinical and radiomics features were extracted, and feature selection was performed. Clinical, radiomics, and combined features were included to construct predictive models. We applied the concept of deep integration to the joint analysis of multiple deep learning methods, used a metaheuristic algorithm to improve the parameter search efficiency, and finally, developed an acute ischemic stroke (AIS) prognosis prediction method, namely, the optimized ensemble of deep learning (OEDL) method.

Results

Among the clinical features, 17 features passed the correlation check. Among the radiomics features, 19 features were selected. In the comparison of the prediction performance of each method, the OEDL method based on the concept of ensemble optimization had the best classification performance. In the comparison to the predictive performance of each feature, the inclusion of the combined features resulted in better classification performance than that of the clinical and radiomics features. In the comparison to the prediction performance of each balanced method, SMOTEENN, which is based on a hybrid sampling method, achieved the best classification performance than that of the unbalanced, oversampled, and undersampled methods. The OEDL method with combined features and mixed sampling achieved the best classification performance, with 97.89, 95.74, 94.75, 94.03, and 94.35% for Macro-AUC, ACC, Macro-R, Macro-P, and Macro-F1, respectively, and achieved advanced performance in comparison with that of methods in previous studies.

Conclusion

The OEDL approach proposed herein could effectively achieve improved stroke prognosis prediction performance, the effect of using combined data modeling was significantly better than that of single clinical or radiomics feature models, and the proposed method had a better intervention guidance value. Our approach is beneficial for optimizing the early clinical intervention process and providing the necessary clinical decision support for personalized treatment.

Machine Learning Methods for Pregnancy and Childbirth Risk Management

Machine learning methods enable medical systems to automatically generate data-driven decision support models using real-world data inputs, eliminating the need for explicit rule design. In this research, we investigated the application of machine learning methods in healthcare, specifically focusing on pregnancy and childbirth risks. The timely identification of risk factors during early pregnancy, along with risk management, mitigation, prevention, and adherence management, can significantly reduce adverse perinatal outcomes and complications for both mother and child. Given the existing burden on medical professionals, clinical decision support systems (CDSSs) can play a role in risk management. However, these systems require high-quality decision support models based on validated medical data that are also clinically interpretable. To develop models for predicting childbirth risks and due dates, we conducted a retrospective analysis of electronic health records from the perinatal Center of the Almazov Specialized Medical Center in Saint-Petersburg, Russia. The dataset, which was exported from the medical information system, consisted of structured and semi-structured data, encompassing a total of 73,115 lines for 12,989 female patients. Our proposed approach, which includes a detailed analysis of predictive model performance and interpretability, offers numerous opportunities for decision support in perinatal care provision. The high predictive performance achieved by our models ensures precise support for both individual patient care and overall health organization management.

Machine Learning for Exposure-Response Analysis: Methodological Considerations and Confirmation of Their Importance via Computational Experimentations

Exposure-response (E-R) is a key aspect of pharmacometrics analysis that supports drug dose selection. Currently, there is a lack of understanding of the technical considerations necessary for drawing unbiased estimates from data. Due to recent advances in machine learning (ML) explainability methods, ML has garnered significant interest for causal inference. To this end, we used simulated datasets with known E-R "ground truth" to generate a set of good practices for the development of ML models required to avoid introducing biases when performing causal inference. These practices include the use of causal diagrams to enable the careful consideration of model variables by which to obtain desired E-R relationship insights, keeping a strict separation of data for model-training and for inference generation to avoid biases, hyperparameter tuning to improve the reliability of models, and estimating proper confidence intervals around inferences using a bootstrap sampling with replacement strategy. We computationally confirm the benefits of the proposed ML workflow by using a simulated dataset with nonlinear and non-monotonic exposure-response relationships.

Deep Learning Methods Applied to Drug Concentration Prediction of Olanzapine

Pharmacometrics and the utilization of population pharmacokinetics play an integral role in model-informed drug discovery and development (MIDD). Recently, there has been a growth in the application of deep learning approaches to aid in areas within MIDD. In this study, a deep learning model, LSTM-ANN, was developed to predict olanzapine drug concentrations from the CATIE study. A total of 1527 olanzapine drug concentrations from 523 individuals along with 11 patient-specific covariates were used in model development. The hyperparameters of the LSTM-ANN model were optimized through a Bayesian optimization algorithm. A population pharmacokinetic model using the NONMEM model was constructed as a reference to compare to the performance of the LSTM-ANN model. The RMSE of the LSTM-ANN model was 29.566 in the validation set, while the RMSE of the NONMEM model was 31.129. Permutation importance revealed that age, sex, and smoking were highly influential covariates in the LSTM-ANN model. The LSTM-ANN model showed potential in the application of drug concentration predictions as it was able to capture the relationships within a sparsely sampled pharmacokinetic dataset and perform comparably to the NONMEM model.

Machine learning algorithms are comparable to conventional regression models in predicting distant metastasis of follicular thyroid carcinoma

OBJECTIVE

Distant metastasis often indicates a poor prognosis, so early screening and diagnosis play a significant role. Our study aims to construct and verify a predictive model based on machine learning (ML) algorithms that can estimate the risk of distant metastasis of newly diagnosed follicular thyroid carcinoma (FTC).

DESIGN

This was a retrospective study based on the Surveillance, Epidemiology, and End Results (SEER) database from 2004 to 2015.

PATIENTS

A total of 5809 FTC patients were included in the data analysis. Among them, there were 214 (3.68%) cases with distant metastasis.

METHOD

Univariate and multivariate logistic regression (LR) analyses were used to determine independent risk factors. Seven commonly used ML algorithms were applied for predictive model construction. We used the area under the receiver-operating characteristic (AUROC) curve to select the best ML algorithm. The optimal model was trained through 10-fold cross-validation and visualized by SHapley Additive exPlanations (SHAP). Finally, we compared it with the traditional LR method.

RESULTS

In terms of predicting distant metastasis, the AUROCs of the seven ML algorithms were 0.746-0.836 in the test set. Among them, the Extreme Gradient Boosting (XGBoost) had the best prediction performance, with an AUROC of 0.836 (95% confidence interval [CI]: 0.775-0.897). After 10-fold cross-validation, its predictive power could reach the best [AUROC: 0.855 (95% CI: 0.803-0.906)], which was slightly higher than the classic binary LR model [AUROC: 0.845 (95% CI: 0.818-0.873)].

CONCLUSIONS

The XGBoost approach was comparable to the conventional LR method for predicting the risk of distant metastasis for FTC.

An introduction to causal inference for pharmacometricians

As formal causal inference begins to play a greater role in disciplines that intersect with pharmacometrics, such as biostatistics, epidemiology, and artificial intelligence/machine learning, pharmacometricians may increasingly benefit from a basic fluency in foundational causal inference concepts. This tutorial seeks to orient pharmacometricians to three such fundamental concepts: potential outcomes, g-formula, and directed acyclic graphs (DAGs).

Application of machine learning techniques to the analysis and prediction of drug pharmacokinetics

In this review, we describe the current status and challenges in applying machine-learning techniques to the analysis and prediction of pharmacokinetic data. The theory of pharmacokinetics has been developed over decades on the basis of physiology and reaction kinetics. Mathematical models allow the reduction of pharmacokinetic data to parameter values, giving insight and understanding into ADME processes and predicting the outcome of different dosing scenarios. However, much information hidden in the data is lost through conceptual simplification with models. It is difficult to use mechanistic models alone to predict diverse pharmacokinetic time profiles, including inter-drug and inter-individual differences, in a cross-sectional manner. Machine learning is a prediction platform that can handle complex phenomena through data-driven analysis. As a resule, machine learning has been successfully adopted in various fields, including image recognition and language processing, and has been used for over two decades in pharmacokinetic research, primarily in the area of quantitative structure-activity relationships for pharmacokinetic parameters. Machine-learning models are generally known to provide better predictive performance than conventional linear models. Owing to the recent success in deep learning, models with new structures are being consistently proposed. These models include transfer learning and generative adversarial networks, which contribute to the effective use of a limited amount of data by diverting existing similar models or generating pseudo-data. How to make such newly emerging machine learning technologies applicable to meet challenges in the pharmacokinetics/pharmacodynamics field is now the key issue.

Value of machine learning algorithms for predicting diabetes risk: A subset analysis from a real-world retrospective cohort study

AIMS/INTRODUCTION

To compare the application value of different machine learning (ML) algorithms for diabetes risk prediction.

MATERIALS AND METHODS

This is a 3-year retrospective cohort study with a total of 3,687 participants being included in the data analysis. Modeling variable screening and predictive model building were carried out using logistic regression (LR) analysis and 10-fold cross-validation, respectively. In total, six different ML algorithms, including random forests, light gradient boosting machine, extreme gradient boosting, adaptive boosting (AdaBoost), multi-layer perceptrons and gaussian naive bayes were used for model construction. Model performance was mainly evaluated by the area under the receiver operating characteristic curve. The best performing ML model was selected for comparison with the traditional LR model and visualized using Shapley additive explanations.

RESULTS

A total of eight risk factors most associated with the development of diabetes were identified by univariate and multivariate LR analysis, and they were visualized in the form of a nomogram. Among the six different ML models, the random forests model had the best predictive performance. After 10-fold cross-validation, its optimal model has an area under the receiver operating characteristic value of 0.855 (95% confidence interval [CI] 0.823-0.886) in the training set and 0.835 (95% CI 0.779-0.892) in the test set. In the traditional LR model, its area under the receiver operating characteristic value is 0.840 (95% CI 0.814-0.866) in the training set and 0.834 (95% CI 0.785-0.884) in the test set.

CONCLUSIONS

In the real-world epidemiological research, the combination of traditional variable screening and ML algorithm to construct a diabetes risk prediction model has satisfactory clinical application value.

Machine learning advances the integration of covariates in population pharmacokinetic models: Valproic acid as an example

Background and Aim: Many studies associated with the combination of machine learning (ML) and pharmacometrics have appeared in recent years. ML can be used as an initial step for fast screening of covariates in population pharmacokinetic (popPK) models. The present study aimed to integrate covariates derived from different popPK models using ML.

Methods: Two published popPK models of valproic acid (VPA) in Chinese epileptic patients were used, where the population parameters were influenced by some covariates. Based on the covariates and a one-compartment model that describes the pharmacokinetics of VPA, a dataset was constructed using Monte Carlo simulation, to develop an XGBoost model to estimate the steady-state concentrations (Css) of VPA. We utilized SHapley Additive exPlanation (SHAP) values to interpret the prediction model, and calculated estimates of VPA exposure in four assumed scenarios involving different combinations of CYP2C19 genotypes and co-administered antiepileptic drugs. To develop an easy-to-use model in the clinic, we built a simplified model by using CYP2C19 genotypes and some noninvasive clinical parameters, and omitting several features that were infrequently measured or whose clinically available values were inaccurate, and verified it on our independent external dataset.

Results: After data preprocessing, the finally generated combined dataset was divided into a derivation cohort and a validation cohort (8:2). The XGBoost model was developed in the derivation cohort and yielded excellent performance in the validation cohort with a mean absolute error of 2.4 mg/L, root-mean-squared error of 3.3 mg/L, mean relative error of 0%, and percentages within ±20% of actual values of 98.85%. The SHAP analysis revealed that daily dose, time, CYP2C19*2 and/or *3 variants, albumin, body weight, single dose, and CYP2C19*1*1 genotype were the top seven confounding factors influencing the Css of VPA. Under the simulated dosage regimen of 500 mg/bid, the VPA exposure in patients who had CYP2C19*2 and/or *3 variants and no carbamazepine, phenytoin, or phenobarbital treatment, was approximately 1.74-fold compared to those with CYP2C19*1/*1 genotype and co-administered carbamazepine + phenytoin + phenobarbital. The feasibility of the simplified model was fully illustrated by its performance in our external dataset.

Conclusion: This study highlighted the bridging role of ML in big data and pharmacometrics, by integrating covariates derived from different popPK models.

Adoption of Machine Learning in Pharmacometrics: An Overview of Recent Implementations and Their Considerations

Pharmacometrics is a multidisciplinary field utilizing mathematical models of physiology, pharmacology, and disease to describe and quantify the interactions between medication and patient. As these models become more and more advanced, the need for advanced data analysis tools grows. Recently, there has been much interest in the adoption of machine learning (ML) algorithms. These algorithms offer strong function approximation capabilities and might reduce the time spent on model development. However, ML tools are not yet an integral part of the pharmacometrics workflow. The goal of this work is to discuss how ML algorithms have been applied in four stages of the pharmacometrics pipeline: data preparation, hypothesis generation, predictive modelling, and model validation. We will also discuss considerations before the use of ML algorithms with respect to each topic. We conclude by summarizing applications that hold potential for adoption by pharmacometricians.

Development and Validation of an Insulin Resistance Model for a Population with Chronic Kidney Disease Using a Machine Learning Approach

Background: Chronic kidney disease (CKD) is a complex syndrome without a definitive treatment. For these patients, insulin resistance (IR) is associated with worse renal and patient outcomes. Until now, no predictive model using machine learning (ML) has been reported on IR in CKD patients. Methods: The CKD population studied was based on results from the National Health and Nutrition Examination Survey (NHANES) of the USA from 1999 to 2012. The homeostasis model assessment of IR (HOMA-IR) was used to assess insulin resistance. We began the model building process via the ML algorithm (random forest (RF), eXtreme Gradient Boosting (XGboost), logistic regression algorithms, and deep neural learning (DNN)). We compared different receiver operating characteristic (ROC) curves from different algorithms. Finally, we used SHAP values (SHapley Additive exPlanations) to explain how the different ML models worked. Results: In this study population, 71,916 participants were enrolled. Finally, we analyzed 1,229 of these participants. Their data were segregated into the IR group (HOMA IR > 3, n = 572) or non-IR group (HOMR IR ≤ 3, n = 657). In the validation group, RF had a higher accuracy (0.77), specificity (0.81), PPV (0.77), and NPV (0.77). In the test group, XGboost had a higher AUC of ROC (0.78). In addition, XGBoost also had a higher accuracy (0.7) and NPV (0.71). RF had a higher accuracy (0.7), specificity (0.78), and PPV (0.7). In the RF algorithm, the body mass index had a much larger impact on IR (0.1654), followed by triglyceride (0.0117), the daily calorie intake (0.0602), blood HDL value (0.0587), and age (0.0446). As for the SHAP value, in the RF algorithm, almost all features were well separated to show a positive or negative association with IR. Conclusion: This was the first study using ML to predict IR in patients with CKD. Our results showed that the RF algorithm had the best AUC of ROC and the best SHAP value differentiation. This was also the first study that included both macronutrients and micronutrients. We concluded that ML algorithms, particularly RF, can help determine risk factors and predict IR in patients with CKD.

Surgical Methods and Social Factors Are Associated With Long-Term Survival in Follicular Thyroid Carcinoma: Construction and Validation of a Prognostic Model Based on Machine Learning Algorithms

Background

This study aimed to establish and verify an effective machine learning (ML) model to predict the prognosis of follicular thyroid cancer (FTC), and compare it with the eighth edition of the American Joint Committee on Cancer (AJCC) model.

Methods

Kaplan-Meier method and Cox regression model were used to analyze the risk factors of cancer-specific survival (CSS). Propensity-score matching (PSM) was used to adjust the confounding factors of different surgeries. Nine different ML algorithms,including eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), Random Forests (RF), Logistic Regression (LR), Adaptive Boosting (AdaBoost), Gaussian Naive Bayes (GaussianNB), K-Nearest Neighbor (KNN), Support Vector Machine (SVM) and Multi-Layer Perceptron (MLP),were used to build prognostic models of FTC.10-fold cross-validation and SHapley Additive exPlanations were used to train and visualize the optimal ML model.The AJCC model was built by multivariate Cox regression and visualized through nomogram. The performance of the XGBoost model and AJCC model was mainly assessed using the area under the receiver operating characteristic (AUROC).

Results

Multivariate Cox regression showed that age, surgical methods, marital status, T classification, N classification and M classification were independent risk factors of CSS. Among different surgeries, the prognosis of one-sided thyroid lobectomy plus isthmectomy (LO plus IO) was the best, followed by total thyroidectomy (hazard ratios: One-sided thyroid LO plus IO, 0.086[95% confidence interval (CI),0.025-0.290], P&lt;0.001; total thyroidectomy (TT), 0.490[95%CI,0.295-0.814], P=0.006). PSM analysis proved that one-sided thyroid LO plus IO, TT, and partial thyroidectomy had no significant differences in long-term prognosis. Our study also revealed that married patients had better prognosis than single, widowed and separated patients (hazard ratios: single, 1.686[95%CI,1.146-2.479], P=0.008; widowed, 1.671[95%CI,1.163-2.402], P=0.006; separated, 4.306[95%CI,2.039-9.093], P&lt;0.001). Among different ML algorithms, the XGBoost model had the best performance, followed by Gaussian NB, RF, LR, MLP, LightGBM, AdaBoost, KNN and SVM. In predicting FTC prognosis, the predictive performance of the XGBoost model was relatively better than the AJCC model (AUROC: 0.886 vs. 0.814).

Conclusion

For high-risk groups, effective surgical methods and well marital status can improve the prognosis of FTC. Compared with the traditional AJCC model, the XGBoost model has relatively better prediction accuracy and clinical usage.

State-of-the-Art Review of Artificial Neural Networks to Predict, Characterize and Optimize Pharmaceutical Formulation

During the development of a pharmaceutical formulation, a powerful tool is needed to extract the key points from the complicated process parameters and material attributes. Artificial neural networks (ANNs), a promising and more flexible modeling technique, can address real intricate questions in a high parallelism and distributed pattern in the manner of biological neural networks. The data mined and analyzing based on ANNs have the ability to replace hundreds of trial and error experiments. ANNs have been used for data analysis by pharmaceutics researchers since the 1990s and it has now become a research method in pharmaceutical science. This review focuses on the latest application progress of ANNs in the prediction, characterization and optimization of pharmaceutical formulation to provide a reference for the further interdisciplinary study of pharmaceutics and ANNs.

The Intratumor Microbiota Signatures Associate With Subtype, Tumor Stage, and Survival Status of Esophageal Carcinoma

Altered human microbiome characteristic has been linked with esophageal carcinoma (ESCA), analysis of microbial profiling directly derived from ESCA tumor tissue is beneficial for studying the microbial functions in tumorigenesis and development of ESCA. In this study, we identified the intratumor microbiome signature and investigated the correlation between microbes and clinical characteristics of patients with ESCA, on the basis of data and information obtained from The Cancer Microbiome Atlas (TCMA) and The Cancer Genome Atlas (TCGA) databases. A total of 82 samples were analyzed for microbial composition at various taxonomic levels, including 40 tumor samples of esophageal squamous cell carcinoma (ESCC), 20 tumor samples of esophageal adenocarcinoma (EAD), and 22 adjacent normal samples. The results showed that the relative abundance of several microbes changed in tumors compared to their paired normal tissues, such as Firmicutes increased significantly while Proteobacteria decreased in tumor samples. We also identified a microbial signature composed of ten microbes that may help in the classification of ESCC and EAD, the two subtypes of ESCA. Correlation analysis demonstrated that compositions of microbes Fusobacteria/Fusobacteriia/Fusobacteriales, Lactobacillales/Lactobacillaceae/Lactobacillus, Clostridia/Clostridiales, Proteobacteria, and Negativicutes were correlated with the clinical characteristics of ESCA patients. In summary, this study supports the feasibility of detecting intratumor microbial composition derived from tumor sequencing data, and it provides novel insights into the roles of microbiota in tumors. Ultimately, as the second genome of human body, microbiome signature analysis may help to add more information to the blueprint of human biology.

Machine Learning Prediction of the Three Main Input Parameters of a Simplified Physiologically Based Pharmacokinetic Model Subsequently Used to Generate Time-Dependent Plasma Concentration Data in Humans after Oral Doses of 212 Disparate Chemicals

Physiologically based pharmacokinetic (PBPK) modeling has the potential to play significant roles in estimating internal chemical exposures. The three major PBPK model input parameters (i.e., absorption rate constants, volumes of the systemic circulation, and hepatic intrinsic clearances) were generated in silico for 212 chemicals using machine learning algorithms. These input parameters were calculated based on sets of between 17 and 65 chemical properties that were generated by in silico prediction tools before being processed by machine learning algorithms. The resulting simplified PBPK models were used to estimate plasma concentrations after virtual oral administrations in humans. The estimated absorption rate constants, volumes of the systemic circulation, and hepatic intrinsic clearance values for the 212 test compounds determined traditionally (i.e., based on fitting to measured concentration profiles) and newly estimated had correlation coefficients of 0.65, 0.68, and 0.77 (p < 0.01, n = 212), respectively. When human plasma concentrations were modeled using traditionally determined input parameters and again using in silico estimated input parameters, the two sets of maximum plasma concentrations (r = 0.85, p < 0.01, n = 212) and areas under the curve (r = 0.80, p < 0.01, n = 212) were correlated. Virtual chemical exposure levels in liver and kidney were also estimated using these simplified PBPK models along with human plasma levels. These results indicate that the PBPK model input parameters for humans of a diverse set of compounds can be reliability estimated using chemical descriptors calculated using in silico tools.

Machine Learning Algorithms to Predict Mortality of Neonates on Mechanical Intubation for Respiratory Failure

BACKGROUND

Early identification of critically ill neonates with poor outcomes can optimize therapeutic strategies. We aimed to examine whether machine learning (ML) methods can improve mortality prediction for neonatal intensive care unit (NICU) patients on intubation for respiratory failure.

METHODS

A total of 1734 neonates with respiratory failure were randomly divided into training (70%, n = 1214) and test (30%, n = 520) sets. The primary outcome was the probability of NICU mortality. The areas under the receiver operating characteristic curves (AUCs) of several ML algorithms were compared with those of the conventional neonatal illness severity scoring systems including the NTISS and SNAPPE-II.

RESULTS

For NICU mortality, the random forest (RF) model showed the highest AUC (0.939 (0.921-0.958)) for the prediction of neonates with respiratory failure, and the bagged classification and regression tree model demonstrated the next best results (0.915 (0.891-0.939)). The AUCs of both models were significantly better than the traditional NTISS (0.836 (0.800-0.871)) and SNAPPE-II scores (0.805 (0.766-0.843)). The superior performances were confirmed by higher accuracy and F1 score and better calibration, and the superior and net benefit was confirmed by decision curve analysis. In addition, Shapley additive explanation (SHAP) values were utilized to explain the RF prediction model.

CONCLUSIONS

Machine learning algorithms increase the accuracy and predictive ability for mortality of neonates with respiratory failure compared with conventional neonatal illness severity scores. The RF model is suitable for clinical use in the NICU, and clinicians can gain insights and have better communication with families in advance.