Ablation Study Results for a Multi-Disease Clinical Prediction Framework

Vijay U. Rathod's study, last updated in May 2026, proposes a unified deep learning framework for disease prediction. The framework was evaluated on three benchmark datasets: the UCI Heart Disease dataset (303 samples), the PIMA Indians Diabetes dataset (768 samples), and a Parkinson's disease voice dataset (195 recordings). The dataset contains the results of this ablation study, showing the incremental impact on mean AUC for different model ensembles.

Use Cases

• Benchmarking deep learning architectures for clinical prediction based on the described framework integrating MLP, CNN, and FT-Transformer.
• Analyzing the impact of dataset characteristics like feature dimensionality and class imbalance on model selection.
• Evaluating ensemble strategies for improving disease prediction performance, as demonstrated with FT-Transformer + autoencoder and CNN + Autoencoder ensembles.
• Studying methods for reliable performance evaluation in clinical ML, based on the described robust preprocessing and nested cross-validation.

Strengths

• Provides specific performance metrics, including an AUC of 0.8980 (±0.0483) for heart disease prediction and 0.8451 (±0.0270) for diabetes classification.
• Framework evaluated on three distinct, publicly available benchmark datasets with known sample sizes (303, 768, and 195).
• Released under a permissive CC-BY-4.0 license, facilitating reuse and modification.

Limitations

• Row count is unknown, which may limit suitability assessment for specific analytical needs.
• Column-level documentation is absent; field semantics must be inferred after download.
• The dataset is very small (5.5 KB), indicating it contains summary results, not the raw clinical data used in the study.

Provenance

Source

figshare, authored by Vijay U. Rathod.

Collection Method

Results from an experimental ablation study of a proposed deep learning framework.

Time Range

Study last updated in 2026; temporal coverage of underlying benchmark datasets is not specified.

Freshness

Last updated 2026-05-08 17:34:40; freshness should be verified.

Geography

Geographic coverage is not specified, but includes the PIMA Indians Diabetes dataset.