Vicomtech – Predictive AI Models for Heart Failure: Readmission and Decompensation Prediction
Vicomtech
Sector: Health
Business Case
Heart failure is a chronic disease responsible for a significant number of hospital admissions, poor quality of life and high hospital costs. The development of predictive models for heart failure both within the in-hospital context for predicting potential readmissions, and in the outpatient environment, with the objective of anticipating potential patient decompensation.
Objectives
Develop a system to predict readmission (e.g., at 30 or 45 days) for heart failure patients upon discharge based on baseline data. In the outpatient setting, attempt to predict decompensation based on telemonitoring data, in order to stratify risk and allow early preventive intervention, avoiding possible hospital admission.
Use case
The in-hospital predictive model is designed to be integrated into the patient’s medical record, so that an alarm would be issued in the event of a clear risk of readmission. The decompensation model is integrated into a telemonitoring system that includes point-of-care devices for measuring certain variables on a daily basis, together with a series of questionnaires. The model is integrated into a decision support system.
Infrastructure
Cloud
Technology
Automatic or Deep Learning
Data
Readmission prediction: baseline data including blood biomarkers and other clinical variables. Decompensation prediction: daily weight, pulse, blood pressure, oxygen saturation, questionnaire.
Resources
1 project manager, 1 AI engineer, supervision from Drs. Vanessa Escolar and Ainara Lozan, clinical validation by the cardiology service of Basurto Hospital. ML models and decision support system developed by Vicomtech. Information system support from the OSI Bilbao-Basurto.
Difficulties and learning
The availability of a well-collected database based on a previous chronicity project greatly facilitated the model development, although in clinical practice this is less common. It is important to properly consider decision-making in order to assess the performance of the model. In this application, a certain number of false positives is acceptable, but if it is excessive the system cannot be scaled.
KPIs (business impact and metrics of the model)
Sensitivity/specificity. Decompensation risk and risk scale. Predictive time window.
Funding
The project was funded by the Department of Health’s health research programme. HAZITEK project: eCardioSurf.
Collaborators, Partners
Cardiology department, Hospital de Basurto, OSI Bilbao-Basurto. BIOEF. Contact with eCardioSurf companies: Ideable, Balidea, STT. Current contact with wearables companies.
