An artificial intelligence (AI)-based approach to post-hoc vital signs processing: study design of the telemedical interventional management in heart failure III (TIM-HF 3) study

Substantial evidence has emerged supporting the use of telemonitoring for patients with heart failure (HF) [1-3]. However, the implementation of new digital care models in healthcare is progressing slowly due to various barriers, such as technological concerns, staff shortages and time constraints [4]. Furthermore, experience from telemedicine studies shows that a considerable proportion of the daily transmitted telemonitoring data contains findings that are clinically unremarkable.

The primary hypothesis of this study is that an AI-based algorithm for daily risk detection will accurately identify HF patients requiring intervention in a post-hoc analysis from a cohort participating in a telemedicine care programme.

The Telemedical Interventional Management in Heart Failure 3 (TIM-HF3) study is a new multicentre cohort study designed to generate data for the retrospective validation of an AI-based algorithm. Telemedical care was provided as part of standard care in accordance with Germany´s quality-assurance agreement for HF telemonitoring. In addition to routine parameters - daily transfer of blood pressure, 2-lead ECG, body weight and self-assessment- participants also measured peripheral oxygen saturation daily and recorded standardised voice samples weekly. A six-minute activity test and a quality of life questionnaire (PROMIS) were carried out during the baseline visit (BV) and the final visit (FV). Follow-up ranged from 6 to 18 months with BV conducted onsite and the FV via telemedicine. Recruitment took place from March 2023 to March 2024. All hospitalisations were adjudicated by an independent Endpoint Committee.

The retrospectively used AI-algorithm, previously described in detail [5], was trained on data of the TIM-HF2 study. To maximize applicability, we applied identical inclusion and exclusion criteria (see picture 1). The algorithm was trained to predict unplanned HF hospitalisations within the following seven days. Baseline risk for each patient was calculated from BV-parameters. Thereafter, a daily risk score indicating the probability of HF hospitalisation within the next seven days was computed from the telemonitoring data. Only patients in the top-risk deciles were flagged for medical assessment by telemedical physicians (see picture 2). From these flags, we derived the sensitivity and specificity for detecting both telemedical interventions and hospitalisations. The goal was to ensure that ≥95% of patients who were hospitalised due to HF were reviewed in the week before the event, despite assessing only ≈30% of patients on any given day.

AI-based algorithms can efficiently filter telemonitoring data to identify patients who need intervention, thereby improving patient care and conserving clinical resources. Nevertheless, integrating AI into routine care demands robust scientific validation and careful implementation planning.

Inclusion and exclusion criteria