Scholarlink

<p>This research offers a framework based on machine learning to achieve real-time physiological monitoring and risk prediction for adverse drug reaction (ADR) surveillance. The proposed pipeline aims to demonstrate through modeling the processing and structuring of continuous biosignal data with a view to using it for early risk scoring in telemetry in low latencies. A hybrid learning strategy is analyzed that consists of a one-dimensional convolutional neural network (1D-CNN) for temporal pattern extraction and Random Forest (RF) classifier for nonlinear decision modeling and feature-level interpretability. They further evaluate an ensemble stacking approach to see if combining the temporal and tree-based predictors improves discrimination performance. By using the PhysioNet Non-EEG physiological dataset which contains multimodal sensor measurements, the computational workflow was validated through experiments. With AUC=0.92, there is no satisfactory class-wise classification accuracy or F1-score in the considered imbalance scenario. In contrast to the previous one, the RF performed in a more stable manner, with an F1-score of 0.61, precision of 0.56, and AUC of 0.96. Performance difference of the stacking ensemble and the RF is rather small indicating tree-based predictors remained the best for this dataset. The results suggest it is possible to deploy machine learning models to produce near real-time physiological risk scoring. It is important to emphasise that the present validation does not use clinically confirmed adverse drug reaction labels or drug-exposure metadata. Instead, the PhysioNet dataset is used as a proxy platform for evaluating the computational workflow, physiological risk scoring capability, and low-latency deployment architecture. Accordingly, the current study should be interpreted as a proof-of-concept framework for real-time physiological monitoring that may support future ADR surveillance after validation on clinically annotated oncology cohorts. However, this is only possible after evaluating the clinical translation of oncology ADR prediction on cohorts with drug-toxicity endpoints.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A machine learning framework for real-time physiological risk scoring toward future adverse drug reaction surveillance

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月3日

查看全文
  • Phatsimo Lobelo,
  • Abid Yahya,
  • Mmoloki Mangwala,
  • Yar Muhammad,
  • Enerst Edozie

摘要

This research offers a framework based on machine learning to achieve real-time physiological monitoring and risk prediction for adverse drug reaction (ADR) surveillance. The proposed pipeline aims to demonstrate through modeling the processing and structuring of continuous biosignal data with a view to using it for early risk scoring in telemetry in low latencies. A hybrid learning strategy is analyzed that consists of a one-dimensional convolutional neural network (1D-CNN) for temporal pattern extraction and Random Forest (RF) classifier for nonlinear decision modeling and feature-level interpretability. They further evaluate an ensemble stacking approach to see if combining the temporal and tree-based predictors improves discrimination performance. By using the PhysioNet Non-EEG physiological dataset which contains multimodal sensor measurements, the computational workflow was validated through experiments. With AUC=0.92, there is no satisfactory class-wise classification accuracy or F1-score in the considered imbalance scenario. In contrast to the previous one, the RF performed in a more stable manner, with an F1-score of 0.61, precision of 0.56, and AUC of 0.96. Performance difference of the stacking ensemble and the RF is rather small indicating tree-based predictors remained the best for this dataset. The results suggest it is possible to deploy machine learning models to produce near real-time physiological risk scoring. It is important to emphasise that the present validation does not use clinically confirmed adverse drug reaction labels or drug-exposure metadata. Instead, the PhysioNet dataset is used as a proxy platform for evaluating the computational workflow, physiological risk scoring capability, and low-latency deployment architecture. Accordingly, the current study should be interpreted as a proof-of-concept framework for real-time physiological monitoring that may support future ADR surveillance after validation on clinically annotated oncology cohorts. However, this is only possible after evaluating the clinical translation of oncology ADR prediction on cohorts with drug-toxicity endpoints.