Human-Centered Design and Iterative Refinement of Tools and Methods to Implement a Surveillance and Risk Prediction System for Clinical Deterioration in Ambulatory Cancer Care

Abstract Background  A common cause of preventable harm is the failure to detect and appropriately respond to clinical deterioration. Timely intervention is needed, particularly in medically complex patients, to mitigate the effects of adverse events, disease progression, and medical error. This challenging problem requires clinical surveillance, early recognition, timely notification of the appropriate clinicians, and effective intervention. Objectives  We determined the feasibility of designing, developing, and implementing the tools and processes to create a surveillance-and-risk prediction system to detect clinical deterioration in cancer outpatients. Methods  We used systems engineering and iterative human-centered design to develop a functional prototype of a surveillance-and-risk prediction system. The system includes passive surveillance involving wearable sensors, active surveillance involving patient event and symptom reporting as well as extraction of selected patient data from the electronic health record (EHR), a predictive model, and communication of estimated risk to clinicians. System usability was evaluated using patient and clinician interviews and clinician ratings using the System Usability Scale (SUS). Results  Fifty of 71 recruited patients enrolled in the feasibility study. Patient-reported outcome measures and clinical data extracted from the EHR were the best predictors of a patient's 7-day risk of experiencing unplanned treatment events (UTEs, i.e., emergency room visits, hospital admissions, or major treatment changes). Deep learning neural network models using these predictors demonstrated modest performance in predicting 7-day UTE risk (PROMS, F-measure: 0.900, area under the receiver operating characteristic curve [AUC-ROC]: 0.983; clinical data from EHR F-measure: 0.625, AUC-ROC: 0.983). Patient risk scores were communicated to clinicians using a risk communication prototype rated favorably by clinicians with a SUS score of 76 out of 100 (median = 80; range: 60–85). Conclusion  We demonstrate the feasibility of a surveillance-and-risk prediction system for detecting and reporting clinical deterioration in cancer outpatients. Future research is needed to fully implement and evaluate system adoption and effectiveness under different clinical situations.