3604 - Detection of Anomalous Patterns in Cancer Patient Undergoing Radiotherapy Using Wearable Sensors: A Proof of Principle Machine Learning Analysis

Purpose/Objective(s): Radiation Therapy (RT) may cause side effects, many of which are subjectively reported by patients. Previous studies showed the possibility to relate subjective symptoms such as fatigue, with data acquired by wearable sensors such as fitness activity trackers (FT) in a breast cancer patients group. Heart Rate (HR) and Activity Level (AL) emerged as good factors to measure fatigue in an objective fashion. On such a basis, our hypothesis is creating a synergy between a machine learning (ML) approach on the data collected via FT and the related Patient Reported Outcome (PRO), it is possible to better define the patient global status of performance during RT. To this intent, the intra-patient HR and AL patterns were analyzed with ML approach, to label them as “regular” or “anomalous.”

Materials/Methods: In this study, a dataset comprising HR and AL data from 20 cancer patients undergoing RT was utilized. Patients were monitored using commercially available FTs to collect the mentioned data, which was sampled at 10-minute intervals over a 50-day period to account for individual daily rhythms. The study applied a data-driven approach to detect deviations from average activity patterns that may indicate physiological distress, e.g., elevated HR while AL is low. The implementation of the Change Point Detection (CPD) algorithm made it possible to identify activity windows for each patient and their subsequent categorization into 3 (i.e., relax, light-activity, and heavy-activity). Subsequently, unsupervised ML models, founded upon Anomaly Detection Algorithms (ADAs) such as Isolation Forest and One-Class SVM, were deployed to identify patient-specific situations. The integration of these data with statistics derived from the sensors’ data allowed to label activity windows within each group as either “regular” or “anomalous.”

Results: Data from one prostate cancer patient was analyzed at first. The ADAs identified “anomalous” patterns on 40% of the samples collected during the monitoring period, 24% of which exhibited a clear mismatch between HR and AL. CPD detected 231 activity windows (80 relax, 144 light-activity, and 7 heavy-activity) within the entire observation period. Subsequent analysis of the activity windows within the selected group revealed that 4% of windows could be labeled as “anomalous” (7 relaxing, 3 light-activity and 0 heavy-activity). Confirmation of labeling validity was obtained via patient-reported questionnaires administered during the monitoring period.

Conclusion: Study’s first results suggest that analyzing wearable sensor data via ML techniques enhances interpretability of the data collected during patient monitoring throughout RT. This allows for more straightforward comprehension of the underlying patterns exhibited by the collected data. Subsequent research activities will center not only on patients’ behavioral analysis, but also on setting a predictive model along with the validation of their clinical applicability across all the population.