Frontiers in digital health · 2026
From behavioral signals to therapy management: a real-world pilot implementation of human-in-the-loop AI workflow in proactive patient care.
СтатьяНазвание и аннотация: Перевод в очереди
Обработка статьи · этапы, ошибки и управление
Загрузка состояния…
Ошибки и ограничения
История · было → стало
Состояние обновляется каждые 5 секунд, пока блок открыт. Процент показывается только там, где известен объём работы. Пауза останавливает продолжение этапа; текущий запрос к источнику может завершиться.
Dickerson David, Lim Koeun, Tourjé Caitlin, Naidu Ramo, Cianni Lindsey, Kibler Andrew B
Аннотация
BACKGROUND: Spinal cord stimulation (SCS) is an established therapy for chronic pain, but post-implant care depends on timely identification of evolving patient needs. Remote monitoring may help detect therapy-use changes that scheduled follow-up and patient-initiated contact can miss. We evaluated Proactive Intelligence, a workflow-integrated, human-in-the-loop AI recommender using passively collected SCS device-interaction data to support care without added patients' data-entry burden.
METHODS: We conducted a prospective, observational real-world pilot among research-consented patients implanted with Prospera SCS devices. Proactive Intelligence used a hybrid neural network incorporating longitudinal therapy-use patterns, operational variables, and patient history/demography to identify patient-days with therapy-adjustment-associated patterns. Model outputs were translated into a prioritized daily review list for the Embrace Care Team (ECT), which reviewed cases, performed outreach when appropriate, adjudicated behavioral-change reasons, and recorded downstream patient care actions. Outcomes included model enrichment, field adjudication yield, intervention distribution, and pain-score change after outreach when follow-up pain scores were available.
RESULTS: Model development used data from 747 permanently implanted patients with sufficient telemetry. Confirmed reprogramming events were rare, occurring at 0.77% per patient-day. At the high-specificity threshold, the model achieved a positive predictive value of 28.8% and sensitivity of 8.6% on a held-out test set, corresponding to approximately 37-fold enrichment over baseline prevalence. During the pilot, predictions were generated across 101,042 patient-days. Across 91 operational days, the ECT reviewed 188 cases involving 175 unique patients. Among reviewed cases, 152 patients responded within three outreach attempts, yielding an 80.9% response rate. Of all reviewed cases, 66.0% were adjudicated as "therapy-related." Among reachable patients, the therapy-related yield was 81.6%, and "Address therapy" actions occurred in 67.8% of cases, including reprogramming and therapy adjustment/discussion. Among cases with follow-up pain scores, therapy-relevant cases showed significant mean NRS reduction after outreach, with improvement after both reprogramming and non-reprogramming therapy adjustment.
DISCUSSION: To our knowledge, this pilot represents the first real-world evaluation of an ECT driven, human-in-the-loop care workflow supported by passively collected SCS device interaction data, without added patient data-entry burden. The findings suggest that machine learning can assist longitudinal SCS care by surfacing behaviorally meaningful changes for timely human review, contextual interpretation, and coordinated support.
METHODS: We conducted a prospective, observational real-world pilot among research-consented patients implanted with Prospera SCS devices. Proactive Intelligence used a hybrid neural network incorporating longitudinal therapy-use patterns, operational variables, and patient history/demography to identify patient-days with therapy-adjustment-associated patterns. Model outputs were translated into a prioritized daily review list for the Embrace Care Team (ECT), which reviewed cases, performed outreach when appropriate, adjudicated behavioral-change reasons, and recorded downstream patient care actions. Outcomes included model enrichment, field adjudication yield, intervention distribution, and pain-score change after outreach when follow-up pain scores were available.
RESULTS: Model development used data from 747 permanently implanted patients with sufficient telemetry. Confirmed reprogramming events were rare, occurring at 0.77% per patient-day. At the high-specificity threshold, the model achieved a positive predictive value of 28.8% and sensitivity of 8.6% on a held-out test set, corresponding to approximately 37-fold enrichment over baseline prevalence. During the pilot, predictions were generated across 101,042 patient-days. Across 91 operational days, the ECT reviewed 188 cases involving 175 unique patients. Among reviewed cases, 152 patients responded within three outreach attempts, yielding an 80.9% response rate. Of all reviewed cases, 66.0% were adjudicated as "therapy-related." Among reachable patients, the therapy-related yield was 81.6%, and "Address therapy" actions occurred in 67.8% of cases, including reprogramming and therapy adjustment/discussion. Among cases with follow-up pain scores, therapy-relevant cases showed significant mean NRS reduction after outreach, with improvement after both reprogramming and non-reprogramming therapy adjustment.
DISCUSSION: To our knowledge, this pilot represents the first real-world evaluation of an ECT driven, human-in-the-loop care workflow supported by passively collected SCS device interaction data, without added patient data-entry burden. The findings suggest that machine learning can assist longitudinal SCS care by surfacing behaviorally meaningful changes for timely human review, contextual interpretation, and coordinated support.