Classifying Adverse Events from SOAP Notes and Sensor Features in a Clinical Trial of Older Adults

Area of Science:

• Geriatric Medicine
• Clinical Trial Safety
• Artificial Intelligence in Healthcare

Background:

• Early detection of adverse events and fall injuries is crucial for patient safety in geriatric clinical trials.
• Rural older adults in clinical trials face unique challenges in monitoring health events.
• Multimodal data integration offers potential for improved patient monitoring.

Purpose of the Study:

• To evaluate multimodal models combining structured SOAP notes and remote biophysical sensor data for classifying adverse events and falls.
• To compare the performance of different machine learning models (XGBoost with various BERT embeddings) with and without sensor features.
• To analyze the impact of cohort-specific documentation practices on model performance.

Main Methods:

• Trained XGBoost classifiers on SOAP note embeddings (BioBERT, BioClinicalBERT, BERT-Uncased) and fused sensor features.
• Compared model performance across control and intervention cohorts for adverse event and fall classification.
• Utilized Area Under the Receiver Operating Characteristic Curve (AUROC) and Recall as primary metrics.
• Analyzed Named Entity Recognition (NER) tokens to understand documentation practice variations.

Main Results:

• Non-fused embeddings showed strong performance for adverse event classification (BioClinicalBERT AUROC=0.89, Recall=0.88 in controls).
• Sensor features excelled in fall classification (control AUROC=0.87, Recall=0.32; intervention AUROC=0.84, Recall=0.12).
• Fusing sensor and embedding features from Subjective and Objective notes achieved near-perfect performance (AUROC=1.0, Recall=1.0).
• Intervention cohort documentation was more patient-focused, correlating with improved event capture.

Conclusions:

• Combining clinical narratives (SOAP notes) with continuous sensor measurements significantly enhances the prediction of adverse events and fall injuries.
• This integrated approach holds promise for increasing clinical trial safety in geriatric populations.
• The findings suggest potential for reducing the frequency of in-person assessments through remote monitoring.