Sermkiat Lolak

Background: Stroke is a major cause of death and disability worldwide. Identifying individuals who would benefit most from preventative interventions, such as antiplatelet therapy, is critical for personalized stroke prevention. However, traditional methods for estimating treatment effects often focus on the average effect across a population and do not account for individual variations in risk and treatment response. Objective: This study aimed to estimate the individualized treatment effects (ITEs) for stroke prevention using a novel combination of Dragonnet, a causal neural network, and conformal inference. The study also aimed to determine and validate the causal effects of known stroke risk factors—hypertension (HT), diabetes mellitus (DM), dyslipidemia (DLP), and atrial fibrillation (AF)—using both a conventional causal model and machine learning models. Methods: A retrospective cohort study was conducted using data from 275,247 high-risk patients treated at Ramathibodi Hospital, Thailand, between 2010 and 2020. Patients aged >18 years with HT, DM, DLP, or AF were eligible. The main outcome was ischemic or hemorrhagic stroke, identified using International Classification of Diseases, 10th Revision (ICD-10) codes. Causal effects of the risk factors were estimated using a range of methods, including: (1) propensity score–based methods, such as stratified propensity scores, inverse probability weighting, and doubly robust estimation; (2) structural causal models; (3) double machine learning; and (4) Dragonnet, a causal neural network, which was used together with weighted split-conformal quantile regression to estimate ITEs. Results: AF, HT, and DM were identified as significant stroke risk factors. Average causal risk effect estimates for these risk factors ranged from 0.075 to 0.097 for AF, 0.017 to 0.025 for HT, and 0.006 to 0.010 for DM, depending on the method used. Dragonnet yielded causal risk ratios of 4.56 for AF, 2.44 for HT, and 1.41 for DM, which is comparable to other causal models and the standard epidemiological case-control study. Mean ITE analysis indicated that several patients with DM or DM with HT, who were not receiving antiplatelet treatment at the time of data collection, showed reductions in total risk of −0.015 and −0.016, respectively. Conclusions: This study provides a comprehensive evaluation of stroke risk factors and demonstrates the feasibility of using Dragonnet and conformal inference to estimate ITEs of antiplatelet therapy for stroke prevention. The mean ITE analysis suggested that those with DM or DM with HT, who were not receiving antiplatelet treatment at the time of data collection, could potentially benefit from this therapy. The findings highlight the potential of these advanced techniques to inform personalized treatment strategies for stroke, enabling clinicians to identify individuals who are most likely to benefit from specific interventions.

Background Machine learning is nowadays commonly used for disease prediction, including cardiovascular disease. There is growing evidence of the effectiveness of machine learning algorithms for stroke risk prediction models. Aims A systematic review was conducted to identify and comprehensively evaluate the available evidence. Summary of review Relevant studies were identified from the three electronic databases (i) MEDLINE via Pubmed, (ii) Scopus, and (iii) IEEE Xplore from inception to 1^st December 2020. Out of 12,626 studies identified, 40 used machine learning for ischemic or hemorrhagic stroke risk prediction models. Synthesis without meta-analysis identified that a boosting algorithm (median C-statistics = 0.9 (interquartile range [IQR]: 0.88-0.92)), and neural network (median C-statistic = 0.80 (IQR: 0.77-0.92)) performed best among ML models in the low risk of bias studies. Moreover, a boosting algorithm also performed best in overall (both low and high risk of bias) studies (median C-statistic = 0.92 (IQR: 0.90-0.95)). Conclusions The systematic review found promising results of the ML algorithm model performances compare with the gold standard conventional models, such as FSRP (C-statistic 0.653) and revised FSRP (C-statistic 0.716). In term of the algorithm, boosting and neural networks are robust, but are considered as black-box models, since they are composed of non-linearity and complex algorithms. It remains questionable whether a physician would adapt these algorithms to use in a real clinical setting. Moreover, less than half of the studies (16 out of 40) were at low risk of bias in our systematic review. More researches with good methodology and study design, alongside explainable and good performance models, may become available in the future. Trial Registration Information The International Prospective Register of Systematic Reviews (PROSPERO) database (ID: CRD42021234081).

Background: Stroke has multiple modifiable and nonmodifiable risk factors and represents a leading cause of death globally. Understanding the complex interplay of stroke risk factors is thus not only a scientific necessity but a critical step toward improving global health outcomes. Objective: We aim to assess the performance of explainable machine learning models in predicting stroke risk factors using real-world cohort data by comparing explainable machine learning models with conventional statistical methods. Methods: This retrospective cohort included high-risk patients from Ramathibodi Hospital in Thailand between January 2010 and December 2020. We compared the performance and explainability of logistic regression (LR), Cox proportional hazard, Bayesian network (BN), tree-augmented Naïve Bayes (TAN), extreme gradient boosting (XGBoost), and explainable boosting machine (EBM) models. We used multiple imputation by chained equations for missing data and discretized continuous variables as needed. Models were evaluated using C-statistics and F1-scores. Results: Out of 275,247 high-risk patients, 9659 (3.5%) experienced a stroke. XGBoost demonstrated the highest performance with a C-statistic of 0.89 and an F1-score of 0.80 followed by EBM and TAN with C-statistics of 0.87 and 0.83, respectively; LR and BN had similar C-statistics of 0.80. Significant factors associated with stroke included atrial fibrillation (AF), hypertension (HT), antiplatelets, HDL, and age. AF, HT, and antihypertensive medication were common significant factors across most models, with AF being the strongest factor in LR, XGBoost, BN, and TAN models. Conclusions: Our study developed stroke prediction models to identify crucial predictive factors such as AF, HT, or systolic blood pressure or antihypertensive medication, anticoagulant medication, HDL, age, and statin use in high-risk patients. The explainable XGBoost was the best model in predicting stroke risk, followed by EBM.

BACKGROUND Stroke has multiple modifiable and nonmodifiable risk factors and represents a leading cause of death globally. Understanding the complex interplay of stroke risk factors is thus not only a scientific necessity but a critical step toward improving global health outcomes. OBJECTIVE We aim to assess the performance of explainable machine learning models in predicting stroke risk factors using real-world cohort data by comparing explainable machine learning models with conventional statistical methods. METHODS This retrospective cohort included high-risk patients from Ramathibodi Hospital in Thailand between January 2010 and December 2020. We compared the performance and explainability of logistic regression (LR), Cox proportional hazard, Bayesian network (BN), tree-augmented Naïve Bayes (TAN), extreme gradient boosting (XGBoost), and explainable boosting machine (EBM) models. We used multiple imputation by chained equations for missing data and discretized continuous variables as needed. Models were evaluated using C-statistics and F1-scores. RESULTS Out of 275,247 high-risk patients, 9659 (3.5%) experienced a stroke. XGBoost demonstrated the highest performance with a C-statistic of 0.89 and an F1-score of 0.80 followed by EBM and TAN with C-statistics of 0.87 and 0.83, respectively; LR and BN had similar C-statistics of 0.80. Significant factors associated with stroke included atrial fibrillation (AF), hypertension (HT), antiplatelets, HDL, and age. AF, HT, and antihypertensive medication were common significant factors across most models, with AF being the strongest factor in LR, XGBoost, BN, and TAN models. CONCLUSIONS Our study developed stroke prediction models to identify crucial predictive factors such as AF, HT, or systolic blood pressure or antihypertensive medication, anticoagulant medication, HDL, age, and statin use in high-risk patients. The explainable XGBoost was the best model in predicting stroke risk, followed by EBM. CLINICALTRIAL

BACKGROUND Machine learning is nowadays commonly used for disease prediction, including cardiovascular disease. There is growing evidence of the effectiveness of machine learning algorithms for stroke risk prediction models. OBJECTIVE A systematic review was conducted to identify and comprehensively evaluate the available evidence. METHODS Relevant studies were identified from the three electronic databases (i) MEDLINE via Pubmed, (ii) Scopus, and (iii) IEEE Xplore from inception to 1st December 2020. Out of 12,626 studies identified, 40 used machine learning for ischemic or hemorrhagic stroke risk prediction models. RESULTS Synthesis without meta-analysis identified that a boosting algorithm (median C-statistics = 0.9 (interquartile range [IQR]: 0.88-0.92)), and neural network (median C-statistic = 0.80 (IQR: 0.79-0.86)) performed best among ML models in the low risk of bias studies. Conversely, the support vector machine performed best in overall (both low and high risk of bias) studies (median C-statistic = 0.89 (IQR: 0.81-0.97)). CONCLUSIONS The systematic review found promising results of the ML algorithm model performances compare with the gold standard conventional models, such as FSRP (C-statistic 0.653) and revised FSRP (C-statistic 0.716). In term of the algorithm, boosting and neural networks are robust, but are considered as black-box models, since they are composed of non-linearity and complex algorithms. It remains questionable whether a physician would adapt these algorithms to use in a real clinical setting. Moreover, less than half of the studies (16 out of 40) were at low risk of bias in our systematic review. More researches with good methodology and study design, alongside explainable and good performance models, may become available in the future. CLINICALTRIAL The International Prospective Register of Systematic Reviews (PROSPERO) database (ID: CRD42021234081).

In the past few years, there are several researches on Parkinson’s disease (PD) recognition using single-photon emission computed tomography (SPECT) images with deep learning (DL) approach. However, the DL model’s complexity usually results in difficultmodel interpretation when used in clinical. Even though there are multiple interpretation methods available for the DL model, there is no evidence of which method is suitable for PD recognition application. This tutorial aims to demonstrate the procedure to choose a suitable interpretationmethod for the PD recogni-tion model. We exhibit four DCNN architectures as an example and introduce six well-known interpretationmethods. Finally, we propose an evaluation method to measure the interpretation performance and a method to use the interpreted feedback for assisting in model selection. The evaluation demonstrates that the guided backpropagation and SHAP interpretation methods are suitable for PD recognition methods in different aspects. Guided backpropagation has the best ability to show fine-grained importance, which is proven by the highest Dice coefficient and lowest mean square error. On the other hand, SHAP can generate a better quality heatmap at the uptake depletion location, which outperforms other methods in discriminating the difference between PD and NC subjects. Shortly, the introduced interpretationmethods can contribute to not only the PD recognition application but also to sensor data processing in an AI Era (interpretable-AI) as feedback in constructing well-suited deep learning architectures for specific applications.

Background: C-reactive protein (CRP) is an inflammatory marker believed to be of value in the early detection of meningitis. We evaluated its potential as a marker for prediction of shunt-related infection in high-risk subjects. Methods: We conducted a prospective pilot study in 26 ventriculoperitoneal shunt procedures; 18 of the patients were considered to be at high risk of infection at the time of shunt insertion. All patients were screened for other disease that could cause, an increase in CRP. Results: The serum CRP medians were 3.90 mg/L in the whole sample and 5.36 mg/L in the high-risk participants. All four shunt infections occurred in the high-risk group (22.2% of the group), three (75%) of which were in patients with meningitis. The logistic regression model showed that CRP concentrations above the cut-off value of ≥7 mg/L were related to shunt infection (p=0.042). The receiving-operating characteristic curve revealed a cutoff point at ≥10.1 mg/L (sensitivity 0.75, 1 – specificity 0.18). The calculated area under the curve was 0.744. The sensitivity and specificity in the whole sample and high-risk group were not different (75% and 79%–80%, respectively). The positive post-test probability was 40% in the whole sample and 50% in the high-risk group. The negative post-test probability was 5% and 7%, respectively. Conclusion: Our data suggest that in a patient at high risk of shunt-related infection, the serum CRP concentration can be a valuable predictor of the risk of infection. Further studies in larger samples would be worthwhile.