Real-world implementation of an eHealth system based on an artificial intelligence designed to predict and reduce emergency department visits by older adults: pragmatic trial.

BACKGROUND: Frail older people use emergency services extensively and digital systems that monitor health remotely could be useful in reducing this use by detecting worsening health conditions earlier.OBJECTIVE: we aimed to implement a system that produces alerts when the machine learning algorithm identifies a short-term risk for emergency department visit and to examine health interventions realized after these alerts and users' experience. This study highlights the feasibility of the general system and its performance in reducing to reduce emergency department visits. It also allowed to evaluate the accuracy alerts' prediction.METHODS: Uncontrolled multi-center trial conducted in community-dwelling older adults receiving assistance from home aides (HAs). We implemented an eHealth system (eHS) that produces an alert for high risk of emergency department visits (EDV). After each home visit, HAs completed a questionnaire on the participants' functional status using a smartphone application and the information was processed in real time by a previously developed machine learning algorithm that identifies patients at risk of EDV within 14 days. In case of risk, the eHS alerted a coordinating nurse who could inform the family carer and the patient's nurses or general practitioner. The primary outcomes were the rate of emergency department visit and number of death after alert-triggered health interventions (ATHI) and users' experience of eHS. Secondary outcome was the accuracy of eHS in predicting EDV.RESULTS: We included 206 patients (average 85 years old (SD 8); 161/206, 78% female) which received aid from 109 HAs, and mean follow-up was 10 months. HAs monitored 2 656 visits which resulted in 405 alerts. Two EDVs were recorded following the 131 alerts with ATHI (2/131, 1.5%), whereas 36 EDVs were recorded following the 274 alerts that did not result in ATHI (36/274, 13.4%) corresponding to an odds-ratio of 0.10 (95% IC 0.02-0.43, p<.0001). Five patients died during the study. All had alerts, 4 did not have ATHI and were hospitalized and 1 had an ATHI (P=.0.038). In terms of overall usability, the digital system was easy to use for 90% (98/109) of HAs and response time was acceptable for 89% (98/109). The sensitivity and specificity of alerts for predicting EDV that occurred within 14 days following the alerts were 83% (95%CI 72-94) and 86% (95%CI 85-87) respectively. The positive and negative predictive values were 9.4% (95%CI 6.5-12.2) and 99.6% (95%CI 99.3-99.9), respectively and positive and negative likelihood ratios were 5.87 (95%CI 4.99-6.92) and 0.20 (95%CI 0.11-0.38), respectively.CONCLUSIONS: The eHS has been successfully implemented, was appreciated by users, and produced relevant alerts. ATHI were associated with a lower rate of EDV, that suggests that the eHS might be useful to lessen EDV in this population.CLINICALTRIAL: clinicaltrials.gov NCT05221697, https://clinicaltrials.gov/ct2/show/NCT05221697.