Applying Machine Learning to Carotid Sonographic Features for Recurrent Stroke in Patients With Acute Stroke

Prognosticating global functional outcome in the recurrent ischemic stroke using baseline clinical and pre-clinical features: A machine learning study

BACKGROUND AND PURPOSE

Recurrent ischemic stroke (RIS) induces additional functional limitations in patients. Prognosticating globally functional outcome (GFO) in RIS patients is thereby important to plan a suitable rehabilitation programme. This study sought to investigate the ability of baseline features for classifying the patients with and without improving GFO (task 1) and identifying patients with poor GFO (task 2) at the third month after discharging from RIS.

METHODS

A total of 86 RIS patients were recruited and divided into the training set and testing set (50:50). The clinical and pre-clinical data were recorded. The outcome was the changes in Modified Rankin Scale (mRS) (task 1) and the mRS score at the third month (mRS 0-2: good GFO, mRS >2: poor GFO) (task 2). The permutation importance ranking method selected features. Four algorithms were trained on the training set with five-fold cross-validation. The best model was tested on the testing set.

RESULTS

In task 1, the support vector machine (SVM) model outperformed the other models, with the high performance matrix on the training set (sensitivity = 0.80; specificity = 1.00) and the testing set (sensitivity = 0.80; specificity = 0.95). In task 2, the SVM model with selected features also performed well on both datasets (training set: sensitivity = 0.76; specificity = 0.92; testing set: sensitivity = 0.72; specificity = 0.88).

CONCLUSION

A machine learning model could be used to classify GFO responses to treatment and identify the third-month poor GFO in RIS patients, supporting physicians in clinical practice.

The performance of machine learning for predicting the recurrent stroke: a systematic review and meta-analysis on 24,350 patients

BACKGROUND

Stroke is a leading cause of death and disability worldwide. Approximately one-third of patients with stroke experienced a second stroke. This study investigates the predictive value of machine learning (ML) algorithms for recurrent stroke.

METHOD

This study was prepared according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guideline. PubMed, Scopus, Embase, and Web of Science (WOS) were searched until January 1, 2024. The quality assessment of studies was conducted using the QUADAS-2 tool. The diagnostic meta-analysis was conducted to calculate the pooled sensitivity, specificity, diagnostic accuracy, positive and negative diagnostic likelihood ratio (DLR), diagnostic accuracy, diagnostic odds ratio (DOR), and area under of the curve (AUC) by the MIDAS package in STATA V.17.

RESULTS

Twelve studies, comprising 24,350 individuals, were included. The meta-analysis revealed a sensitivity of 71% (95% CI 0.64-0.78) and a specificity of 88% (95% confidence interval (CI) 0.76-0.95). Positive and negative DLR were 5.93 (95% CI 3.05-11.55) and 0.33 (95% CI 0.28-0.39), respectively. The diagnostic accuracy and DOR was 2.89 (95% CI 2.32-3.46) and 18.04 (95% CI 10.21-31.87), respectively. The summary ROC curve indicated an AUC of 0.82 (95% CI 0.78-0.85).

CONCLUSION

ML demonstrates promise in predicting recurrent strokes, with moderate to high sensitivity and specificity. However, the high heterogeneity observed underscores the need for standardized approaches and further research to enhance the reliability and generalizability of these models. ML-based recurrent stroke prediction can potentially augment clinical decision-making and improve patient outcomes by identifying high-risk patients.

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.

A deep learning algorithm to identify carotid plaques and assess their stability

Background

Carotid plaques are major risk factors for stroke. Carotid ultrasound can help to assess the risk and incidence rate of stroke. However, large-scale carotid artery screening is time-consuming and laborious, the diagnostic results inevitably involve the subjectivity of the diagnostician to a certain extent. Deep learning demonstrates the ability to solve the aforementioned challenges. Thus, we attempted to develop an automated algorithm to provide a more consistent and objective diagnostic method and to identify the presence and stability of carotid plaques using deep learning.

Methods

A total of 3,860 ultrasound images from 1,339 participants who underwent carotid plaque assessment between January 2021 and March 2023 at the Shanghai Eighth People's Hospital were divided into a 4:1 ratio for training and internal testing. The external test included 1,564 ultrasound images from 674 participants who underwent carotid plaque assessment between January 2022 and May 2023 at Xinhua Hospital affiliated with Dalian University. Deep learning algorithms, based on the fusion of a bilinear convolutional neural network with a residual neural network (BCNN-ResNet), were used for modeling to detect carotid plaques and assess plaque stability. We chose AUC as the main evaluation index, along with accuracy, sensitivity, and specificity as auxiliary evaluation indices.

Results

Modeling for detecting carotid plaques involved training and internal testing on 1,291 ultrasound images, with 617 images showing plaques and 674 without plaques. The external test comprised 470 ultrasound images, including 321 images with plaques and 149 without. Modeling for assessing plaque stability involved training and internal testing on 764 ultrasound images, consisting of 494 images with unstable plaques and 270 with stable plaques. The external test was composed of 279 ultrasound images, including 197 images with unstable plaques and 82 with stable plaques. For the task of identifying the presence of carotid plaques, our model achieved an AUC of 0.989 (95% CI: 0.840, 0.998) with a sensitivity of 93.2% and a specificity of 99.21% on the internal test. On the external test, the AUC was 0.951 (95% CI: 0.962, 0.939) with a sensitivity of 95.3% and a specificity of 82.24%. For the task of identifying the stability of carotid plaques, our model achieved an AUC of 0.896 (95% CI: 0.865, 0.922) on the internal test with a sensitivity of 81.63% and a specificity of 87.27%. On the external test, the AUC was 0.854 (95% CI: 0.889, 0.830) with a sensitivity of 68.52% and a specificity of 89.49%.

Conclusion

Deep learning using BCNN-ResNet algorithms based on routine ultrasound images could be useful for detecting carotid plaques and assessing plaque instability.

Efficacy and safety of Shenmai injection for acute ischemic stroke: a systematic review and meta-analysis

Background

Ischemic stroke is a serious and sudden cerebrovascular condition that significantly affects individual's health and imposes a substantial economic burden on medical management. Despite its widespread use in China, there is still a lack of reliable evidence regarding the efficacy of Shenmai injection (SMI) in acute ischemic stroke (AIS). We aimed to comprehensively assess the effectiveness and safety of SMI in treating AIS through a systematic review and meta-analysis.

Methods

Randomized controlled studies (RCTs) investigating the efficacy of SMI in treating AIS were searched for in eight databases from the inception of each database till January 2024. We utilized the ROB 2.0 to assess the risk of bias. A meta-analysis was conducted using Review Manager 5.4, while sensitivity analyses and publication bias assessments were conducted using Stata 16.1.

Results

A total of 17 studies involving 1,603 AIS patients were included in our meta-analysis. Our results showed that SMI plus conventional treatments (CTs) was more effective than CTs alone in improving the total effective rate (RR 1.22, 95% CI: 1.14 to 1.30, p < 0.00001), the Barthel index (BI) (MD 12.18, 95% CI: 10.30 to 14.06, p < 0.00001), and reducing the National Institute of Health Stroke Scale Score (NIHSS) score (MD -3.05, 95% CI: 3.85 to -2.24, p < 0.00001) and Modified Rankin Scale (mRS) (MD -0.68, 95% CI: 0.86 to-0.49, p < 0.00001). In addition, SMI combination therapy was better than CTs alone in decreasing the levels of IL-6, IL-18, and hs-CRP. SMI therapy also enhanced the cerebral hemorheology of patients by reducing levels of fibrinogen and plasma viscosity. However, there was no significant difference in the incidence of adverse events, including elevated transaminase, rash, nausea, bleeding, urticaria, headache, vomiting, chest tightness, and facial flushes. Moreover, no serious adverse effects or life-threatening events were reported.

Conclusion

Our study shows that combining SMI with CTs effectively enhances the neurological function of patients with acute cerebral infarction. However, our findings should be interpreted considering the significant heterogeneity and suboptimal quality of the analyzed trials.

Systematic review registration

https://www.crd.york.ac.uk/prospero/display_record.php?ID=CRD42024504675, Identifier PROSPERO, CRD42024504675.

Interpretable machine learning model integrating clinical and elastosonographic features to detect renal fibrosis in Asian patients with chronic kidney disease

BACKGROUND

Non-invasive renal fibrosis assessment is critical for tailoring personalized decision-making and managing follow-up in patients with chronic kidney disease (CKD). We aimed to exploit machine learning algorithms using clinical and elastosonographic features to distinguish moderate-severe fibrosis from mild fibrosis among CKD patients.

METHODS

A total of 162 patients with CKD who underwent shear wave elastography examinations and renal biopsies at our institution were prospectively enrolled. Four classifiers using machine learning algorithms, including eXtreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), Light Gradient Boosting Machine (LightGBM), and K-Nearest Neighbor (KNN), which integrated elastosonographic features and clinical characteristics, were established to differentiate moderate-severe renal fibrosis from mild forms. The area under the receiver operating characteristic curve (AUC) and average precision were employed to compare the performance of constructed models, and the SHapley Additive exPlanations (SHAP) strategy was used to visualize and interpret the model output.

RESULTS

The XGBoost model outperformed the other developed machine learning models, demonstrating optimal diagnostic performance in both the primary (AUC = 0.97, 95% confidence level (CI) 0.94-0.99; average precision = 0.97, 95% CI 0.97-0.98) and five-fold cross-validation (AUC = 0.85, 95% CI 0.73-0.98; average precision = 0.90, 95% CI 0.86-0.93) datasets. The SHAP approach provided visual interpretation for XGBoost, highlighting the features' impact on the diagnostic process, wherein the estimated glomerular filtration rate provided the largest contribution to the model output, followed by the elastic modulus, then renal length, renal resistive index, and hypertension.

CONCLUSION

This study proposed an XGBoost model for distinguishing moderate-severe renal fibrosis from mild forms in CKD patients, which could be used to assist clinicians in decision-making and follow-up strategies. Moreover, the SHAP algorithm makes it feasible to visualize and interpret the feature processing and diagnostic processes of the model output.

Development and Validation of Data-Level Innovation Data-Balancing Machine Learning Models for Predicting Optimal Implantable Collamer Lens Size and Postoperative Vault

INTRODUCTION

There are only four sizes of implantable collamer lens (ICL) available for selection, which cannot completely fit all patients as a result of the discontinuity of ICL sizes. Sizing an optimal ICL and predicting postoperative vault are still unresolved problems. This study aimed to develop and validate innovative data-level data-balancing machine learning-based models for predicting ICL size and postoperative vault.

METHODS

The patients were randomly assigned to training and internal validation sets in a 4:1 ratio. Feature selection was performed using analysis of variance (ANOVA) and Kruskal-Wallis feature importance methods. Traditional linear regression model and machine learning-based models were used. The accuracy of models was assessed using the area under the curve (AUC) and confusion matrix.

RESULTS

A total of 564 patients (1127 eyes) were eligible for this study, consisting of 808 eyes in the training set, 202 eyes in the internal validation set, and 117 eyes in the external validation set. Compared with the traditional linear regression method, the machine learning model bagging tree showed the best performance for ICL size selection, with an accuracy of 84.5% (95% confidence interval (CI) 83.2-85.8%), and the AUC ranged from 0.88 to 0.99; the prediction accuracy of 12.1 mm and 13.7 mm ICL sizes was improved by 49% and 59%, respectively. The bagging tree model achieved the best accuracy [90.2%, (95% CI 88.9-91.5%)] for predicting the postoperative vault, and the AUC ranged from 0.90 to 0.94. The prediction accuracies of internal and external validation dataset for ICL sizing were 82.2% (95% CI 81.1-83.3%) and 82.1% (95% CI 81.1-83.1%), respectively.

CONCLUSIONS

The innovative data-level data balancing-based machine learning model can be used to predict ICL size and postoperative vault more accurately, which can assist surgeons in choosing optimal ICL size, thus reducing risks of postoperative complications and secondary surgery.

Artificial intelligence-based model for dose prediction of sertraline in adolescents: a real-world study

BACKGROUND

Variability exists in sertraline pharmacokinetic parameters in individuals, especially obvious in adolescents. We aimed to establish an individualized dosing model of sertraline for adolescents with depression based on artificial intelligence (AI) techniques.

METHODS

Data were collected from 258 adolescent patients treated at the First Hospital of Hebei Medical University between December 2019 to July 2022. Nine different algorithms were used for modeling to compare the prediction abilities on sertraline daily dose, including XGBoost, LGBM, CatBoost, GBDT, SVM, ANN, TabNet, KNN, and DT. Performance of four dose subgroups (50 mg, 100 mg, 150 mg, and 200 mg) were analyzed.

RESULTS

CatBoost was chosen to establish the individualized medication model with the best performance. Six important variables were found to be correlated with sertraline dose, including plasma concentration, PLT, MPV, GL, A/G, and LDH. The ROC curve and confusion matrix exhibited the good prediction performance of CatBoost model in four dose subgroups (the AUC of 50 mg, 100 mg, 150 mg, and 200 mg were 0.93, 0.81, 0.93, and 0.93, respectively).

CONCLUSION

The AI-based dose prediction model of sertraline in adolescents with depression had a good prediction ability, which provides guidance for clinicians to propose the optimal medication regimen.

Transplant renal artery stenosis: utilization of machine learning to identify ancillary sonographic and doppler parameters to predict stenosis in patients with graft dysfunction

PURPOSE

To determine if ancillary sonographic and Doppler parameters can be used to predict transplant renal artery stenosis in patients with renal graft dysfunction.

MATERIALS AND METHODS

IRB-approved, HIPAA-compliant retrospective study included 80 renal transplant patients who had renal US followed by renal angiogram between January 2018 and December 2019. A consensus read of two radiologists recorded these parameters: peak systolic velocity, persistence of elevated velocity, grayscale narrowing, parvus tardus, delayed systolic upstroke, angle of the systolic peak (SP angle), and aliasing. Univariate analysis using t-test or chi-square was performed to determine differences between patients with and without stenosis. P values under 0.05 were deemed statistically significant. We used machine learning algorithms to determine parameters that could better predict the presence of stenosis. The algorithms included logistic regression, random forest, imbalanced random forest, boosting, and CART. All 80 cases were split between training and testing using stratified sampling using a 75:25 split.

RESULTS

We found a statistically significant difference in grayscale narrowing (p = 0.0010), delayed systolic upstroke (p = 0.0002), SP angle (p = 0.0005), and aliasing (p = 0.0024) between the two groups. No significant difference was found for an elevated peak systolic velocity (p = 0.1684). The imbalanced random forest (IRF) model was selected for improved accuracy, sensitivity, and specificity. Specificity, sensitivity, AUC, and normalized Brier score for the IRF model using all parameters were 73%, 81%, 0.82, and 69 in the training set, and 78%, 58%, 0.78, and 80 in the testing set. VIMP assessment showed that the combination of variables that resulted in the most significant change of the training set performance was that of grayscale narrowing and SP angle.

CONCLUSION

Elevated peak systolic velocity did not discriminate between patients with and without TRAS. Adding ancillary parameters into the machine learning algorithm improved specificity and sensitivity similarly in the training and testing sets. The algorithm identified the combination of lumen narrowing coupled with the angle of the systolic peak as better predictor of TRAS. This model may improve the accuracy of ultrasound for transplant renal artery stenosis.

Fully automatic carotid arterial stiffness assessment from ultrasound videos based on machine learning

Arterial stiffness (AS) refers to the loss of arterial compliance and alterations in vessel wall properties. The study of local carotid stiffness (CS) is particularly important since carotid artery stiffening raises the risk of stroke, cognitive impairment, and dementia. So, stiffness measurement as a screening tool approach is crucial because it can reduce mortality and facilitate therapy planning. This study aims to evaluate the stiffness of the CCA using machine learning (ML) through the features of diameter change (ΔD) and stiffness parameters. This study was conducted in seven stages: data collection, preprocessing, CCA segmentation, CCA lumen diameter (DCCA) computing during cardiac cycles, denoising signals of DCCA, computational of AS parameters, and stiffness assessment using ML. The 51 videos (with 25 s) of CCA B-mode ultrasound (US) were used and analyzed. Each US video yielded approximately 750 sequential frames spanning about 24 cardiac cycles. Firstly, US preset settings with time gain compensation with a U-pattern were employed to enhance CCA segmentations. The study showed that auto region-growing, employed three times, is appropriate for segmenting walls with a short running time (4.56 s/frame). The diameter computed for frames constructs a signal (diameter signal) with noisy parts in the shape of peak variance and an un-smooth side. Among the 12 employed smoothing methods, spline fitting with a mean peak difference per cycle (MPDCY) of 0.58 pixels was the most effective for the diameter signal. The authors propose the MPDCY as a new selection criterion for smoothing methods with highly preserved peaks. The ΔD (D_sys-D_dia) determined in this study was validated by statistical analysis as a viable replacement for manual ΔD measurement. Statistical analysis was carried out by Mann-Whitney t-test with a p-value of 0.81, regression line R² = 0.907, and there was no difference in means between the two groups for box plots. The stiffness parameters of the carotid arteries were calculated based on auto-ΔD and pulse pressure. Five ML models, including K-nearest neighbor (KNN), support vector machine (SVM), decision tree (DT), logistic regression (LR), and random forest (RF), fed by distension (ΔD) and five stiffness parameters, were used to distinguish between the stiffened and un-stiffened CCA. Except for SVM, all models performed excellently in terms of specificity, sensitivity, precision, and area under the curve (AUC). In addition, the scatterplot and statistical analysis of the fed features confirm these remarkable outcomes. The scatter plot demonstrates that a linear hyperline can easily distinguish between the two classes. The statistical analysis shows that the stiffness parameters computed from the database of this work were statistically (p < 0.05) distributed into the non-stiffness and stiffness groups. The presented models are validated by applying them to additional datasets. Applying models to other datasets reveals a model performance of 100%. The proposed ML models could be applied in clinical practice to detect CS early, which is essential for preventing stroke.