Your Heart Failure Prediction to Identify Un-diagnosed Patients from Routine Primary Care Records

Read the full article See related articles

Discuss this preprint

Start a discussion What are Sciety discussions?

Listed in

This article is not in any list yet, why not save it to one of your lists.
Log in to save this article

Abstract

Heart Failure is a common and serious condition that often remains undetected until a major cardio-vascular event leads to diagnosis is secondary care. Here we propose a portable artificial intelligence tool that integrates clinical guidelines with phenotypic markers to identify high-risk patients who may benefit from formal diagnosis evaluation and timely initiation of treatment.

Methods

Diagnosis guidelines are first encoded using a rule-based model, which is then used to train a neural network. Relying on de-identified real-world evidence from UK primary care, transfer learning is used to train on 91,346 historical records and forecast the 6.2% patients who received a diagnosis within 3 years. Tested for portability in an independent sample consisting of 56,308 validation records, predictions are interpreted using Shapley values and individually assessed for statistical significance by comparison with matched digital twin cohorts. A Kaplan-Meier survival analysis links positive predictions to the observed excess mortality.

Results

Compared with the prevailing challenge of under-diagnosis, model predictions in the validation set (0.7% TP, 2.7% FP) demonstrate strong statistical support, with fewer than 1.5% failing to reject a null hypothesis at p=0.05. Among the TP, the likelihood of receiving a future diagnosis is over 7.6 times higher than the baseline prevalence in the validation cohort. In both TP and FP cohorts, patients aged 60-70 years exhibited mortality rates more than fivefold higher than the control population. Furthermore, variables derived from the Complete Blood Count (CBC) including white blood cell count (WBC) and red cell distribution width (RDW), contribute significant predictive value beyond established diagnosis criteria.

Conclusions

When implemented within a clinical decision support system, predictive AI has the potential to improve patients outcomes by leveraging routinely collected phenotypic markers, which are challenging for clinicians to interpret in the context of complex decision-making pathways. 1

Article activity feed