TITLE:
A Bayesian Neural Network Prediction of Type 1 Diabetes Mellitus on the TEDDY Cohort
AUTHORS:
Mwagandi Kalume Kiponda, Aggrey Onago Adem, Leonard Kiti Alii, Michael Mbindyo Munywoki, Lazarus K. Mramba, Martin K. Haro, Josephat Mutembei
KEYWORDS:
Bayesian Neural Network, Type 1 Diabetes Mellitus, Class Imbalance, Semiparametric Modelling, Weighted ELBO, Monte Carlo Dropout, Epistemic Uncertainty, Accuracy Paradox, TEDDY Cohort, Paediatric Screening
JOURNAL NAME:
Journal of Biosciences and Medicines,
Vol.14 No.10,
October
23,
2026
ABSTRACT: Predicting the onset of Type 1 Diabetes Mellitus (T1DM) in genetically at-risk children presents a fundamental statistical challenge: disease prevalence in prospective cohorts rarely exceeds 4% - 5%, producing severe class imbalance that causes standard classifiers to optimise for the majority class while catastrophically failing the minority. We address this through a semiparametric Bayesian Neural Network (BNN) evaluated on 8556 children from the multicentre TEDDY cohort (331 T1DM cases; 3.9% prevalence). The BNN integrates a Gaussian-prior regularised feedforward architecture with Bayesian weight distributions, optimised via a Weighted Surrogate Evidence Lower Bound (W-ELBO) that embeds class-specific cost penalties directly into the variational objective. Uncertainty was quantified through Monte Carlo Dropout with
T=50
stochastic forward passes. In a prespecified 6 × 3 factorial design comparing six model families under three imbalance-mitigation conditions, the BNN trained with raw data and inverse-frequency class weights achieved the highest F1 score (0.6426) and Matthews Correlation Coefficient (MCC = 0.6401), with Sensitivity of 80.81% (95% CI: 71.66% - 88.03%) and Specificity of 97.16% (95% CI: 96.43% - 97.78%); calibration was formally assessed via the Brier score (0.0510) together with Expected/Maximum Calibration Error and Cox calibration statistics (§4.5), and quantified epistemic uncertainty (mean predictive SD = 0.0671) — capabilities absent in all deterministic comparators. SMOTE oversampling was found to be redundant when combined with class-weighted loss (producing mathematically identical results to four decimal places) and actively harmful for kernel-based SVMs (0% Sensitivity at the default 0.5 threshold despite AUC = 0.8644), constituting a novel methodological finding. These results support a paradigm shift from accuracy-optimised deterministic pipelines to sensitivity-prioritised, uncertainty-aware semiparametric frameworks for low-prevalence clinical prediction.