Abstract Introduction Heart failure (HF) burden on public health is substantial and efficient patient management strategies are needed. Traditional methods of HF monitoring face limitations in scalability and effectiveness. Artificial Intelligence (AI) might offer novel solutions for enhanced resource management and alleviation of healthcare burdens. Purpose To study the feasibility of implementing automatic remote monitoring via phone calls operated by AI in a HF unit. Methods We conducted a non-randomized, open, pilot trial employing an AI technology based on natural language processing for autonomous telephone outreach for the follow up of HF patients. 86 patients from a tertiary hospital HF unit undergoing pharmacological titration and education were enrolled. The AI conducted weekly calls to gather symptomatology, vital signs, and health status changes, which were then displayed on a virtual platform featuring alerts preconfigured by the research team. These alerts guided patient communication and intervention as deemed necessary by a specialized nurse. Results Characteristics of patients are shown in Table 1. The AI followed these patients for a median of 251.5 days. 2773 telephonic calls were performed. We estimate an equivalent to 1445h of work (17.2 days/month). Adherence was high during the study, with 98% of the calls completed. 1427 alerts were generated, being the most frequent weight change (251; 17.6%), orthopnea (220; 15.4%) and fatigue (187; 13.1%). It led to 487 remote interventions, most of them consisting in calling the patient for assessment (409; 84.0%), but also in adjusting medication remotely (34; 7.0%). See Figure 2. A HF telemedicine specialized nurse required 205min per week for alert review, 2.38min/patient-week. 2 patients died, one from cardiovascular causes, and 7 patients had a HF hospitalization, 42.9% with previous alarms. 20 events of worsening HF were recorded, comprising hospitalization, emergency visit or IV diuretic administration. Oral diuretic dose was increased in 22 occasions, 15 of them remotely in response to an alert raised by the AI operator. At 6 months functional class globally improved (p = 0.028), but NTproBNP increased (p=0.0472). Quality of life was evaluated at baseline and at 6 months, with no change in EQ5D5L (median 0.717, change -0.005, p 0.71) or PHQ4 (median 2.88, change -0.29, p 0.7792), but KCCQ12 total score worsened (median 74.49, change -2.32, p 0.048). Experience was rated with 8.97/10 in a satisfaction questionnaire. All patients recommended the program, remarking a closer follow up, better communication and a safety feeling. Conclusions AI enables closer monitoring of patients with HF, maintaining contact with stable patients while also allowing for resource allocation and tracking of higher-risk patients or those experiencing changes in symptoms. A randomized clinical trial comparing to standard follow-up is needed to evaluate this promising technology.Table 1Figure 2
Read full abstract