• Home  
  • Extra Clinical Features Failed to Improve Heart AI in Small Benchmark
- Research Briefs

Extra Clinical Features Failed to Improve Heart AI in Small Benchmark

Adding eight engineered indicators failed to consistently improve heart disease prediction in a controlled 270-record benchmark.

Adding eight clinically motivated indicators did not consistently improve a heart-disease prediction model in a small benchmark. The results caution against assuming that extra features automatically produce better medical AI.

What researchers studied

Researchers analysed the UCI Statlog Heart dataset of 270 records. One neural network used 13 original clinical predictors, while another included eight additional deterministic indicators derived from the same inputs. A fixed stratified 80:20 split produced 216 training and 54 held-out test records. Five-fold cross-validation was confined to the training portion. The researchers compared several classification metrics rather than relying on accuracy alone.

What they found

On the held-out test set, the original-input model achieved 83.33% accuracy versus 79.63% for the augmented model. The augmented model had a slightly higher ROC-AUC, 0.879 against 0.867, but that isolated difference did not establish an overall improvement. All five cross-validation mean metrics favoured the clinical-only model. The authors also clarified that their experiment evaluated deterministic feature augmentation rather than formal ontology reasoning.

What it means

The results illustrate why model developers should test whether engineered features actually add useful information. Clinically meaningful labels can still duplicate existing inputs, and a single favourable metric may hide weaker performance elsewhere. Transparent partitions, multiple measures and reproducible comparisons are particularly important for healthcare applications. These experimental findings do not establish the effectiveness of either network in real clinical decision-making.

Important limitation

The dataset was small and the test portion contained only 54 records, making estimates sensitive to sampling variation. The original ontology resources were unavailable, and some earlier explanation material could not be validated. No prospective clinical evaluation was performed. Performance on a historical benchmark should not be extrapolated to hospitals or used to recommend patient treatment.

Source Information

Study: A controlled ablation study of clinically motivated engineered features for heart disease prediction on the Statlog dataset

Authors: U. Priyadharshini and R. Vijayan

Journal / Source: Frontiers in Artificial Intelligence

Published: 20 November 2026

DOI: 10.3389/frai.2026.2010348

Contact Us

Research Today is a South African digital publication that makes credible research easier to understand.

TERMS OF USE & PRIVACY POLICY

follow us