Abstract / Summary
Abstract Background Many patients with cancer experience pain; however, expressing their pain levels can be challenging for terminally ill patients. Although machine learning (ML) models for pain detection using biological signals are being developed, studies collecting long-term data on patients are insufficient. Therefore, we aimed to develop an ML model for detecting pain due to cancer using data from wrist-type devices in clinical settings. Methods We recruited patients with a numerical rating scale (NRS; 0–10) score ≥ 4, hospitalized between August 2023 and April 2024. The patients wore the device and recorded the timestamp of pain onset or changes and NRS in real-time. Using a 5-min sliding window, summary statistics were calculated. Two positive labels were assigned: moderate or higher pain (NRS ≥ 4) and pain exceeding the personalized pain goal (PPG). The F1 score was the primary evaluation metric used to evaluate the ML model. Results Nineteen patients were enrolled; one patient was excluded because of inadequate data. The average age was 64.9 years; 61.1% were males, and 61.1% had somatic pain. The lower abdomen was the most common site of pain. For the common model, the average F1 score for detecting moderate or high pain was 0.60, with an accuracy of 0.81. The F1 score for detecting pain exceeding the PPG was 0.44, with an accuracy of 0.67. In a sensitivity analysis excluding static patient characteristics, performance for detecting moderate or higher pain fell substantially (mean F1 0.62 to 0.30; mean AUROC 0.77 to 0.52), whereas the reduction for detecting pain exceeding the PPG was much smaller (F1 0.45 to 0.39; AUROC 0.66 to 0.59). Conclusions This study demonstrates the feasibility of collecting longitudinal wearable signals together with contemporaneous pain reports in patients with cancer, and provides preliminary evidence that these signals carry information associated with pain states. The current performance is insufficient for reliable automated pain detection, and larger cohorts with rigorous patient-level and external validation are required.