Deep learning-based genetic disease prediction with Monte Carlo (MC) dropout uncertainty estimation: a comparative study with classical machine learning baselines

Deep neural networks have been successfully applied to structured clinical data with a good predictive performance, but their point predictions do not reflect an indication of confidence, which limits their trustworthiness in screening for genetic conditions where an overconfident misclassification may delay the appropriate confirmatory testing or intervention.

A feed-forward neural network (128-64-32 hidden units, batch normalization, and dropout p = 0.3) was trained to classify patients into five genetic-disease-associated groups based on twelve regularly accessible clinical and molecular biomarkers, using a publicly available, pre-labelled structured biomedical dataset (n = 1,000) whose real-world clinical vs. synthetic provenance could not be independently confirmed (see Section 2.1 and Limitations). Performance was assessed using stratified 5-fold cross-validation versus four traditional baselines (logistic regression, random forest, SVM, gradient boosting; McNemar's test for pairwise significance). We quantified predictive uncertainty using Monte Carlo (MC) Dropout (T up to 200 stochastic passes), decomposed it into epistemic and aleatoric components, and evaluated it thru calibration analysis (ECE, MCE, Brier score, negative log-likelihood), an uncertainty-quality analysis (low/medium/high uncertainty strata), a selective-prediction (risk-coverage) analysis, a seven-configuration ablation study, and robustness testing across random seeds, reduced training data, and Gaussian feature noise. Results The developed network achieved 94.30% +- 1.36% cross-validated accuracy which was statistically comparable to logistic regression (p = 0.092) and gradient boosting (p = 0.238) but considerably inferior to random forest (95.50% +- 1.26%, p = 0.0018) and SVM (95.30% +- 1.36%, p = 0.0063). MC-Dropout obtained 93.33% accuracy (macro F1 = 0.933, macro ROC-AUC = 0.971, MCC = 0.917) with ECE = 0.076 on a held-out test partition-numerically the highest (worst) ECE among the six model/inference configurations evaluated, exceeding all four classical baselines (ECE 0.024-0.066), although MC-Dropout inference reduced negative log-likelihood substantially relative to deterministic inference (0.370 vs. 0.445), indicating its calibration benefit is concentrated in reducing the penalty for confident errors rather than in bin-wise confidence-accuracy alignment. Errors were about twice as likely in the highest (most uncertain) tercile of predictions than in the lowest (12.0% vs. 6.0% mistake rate), and 60% of all misclassifications were within the high-uncertainty group. Removing the most doubtful 20%-30% of predictions yielded a retained-sample accuracy of above 95%. Accuracy and calibration were consistent across five random seeds (93.07%) and deteriorated gracefully with increasing predictive entropy under synthetic feature-noise perturbation.

On this structured, moderate-sized genetic-disease benchmark dataset-evaluated entirely through internal computational validation, without external or prospective clinical testing-classical machine learning algorithms achieved statistically comparable or slightly higher raw classification accuracy than this deep learning approach under standard, non-tuned hyperparameters for all models.