An explainable hyperparameter-optimized genetic algorithm–backpropagation (GA–BP) neural network model for medical software cost estimation

Background Accurate and transparent cost estimation is pivotal for medical software, where regulatory constraints and integration complexity amplify the risk of misestimation. Traditional models rarely encode domain drivers or provide audit-ready explanations. Methods We propose an explainable, hyperparameter-optimized Genetic Algorithm–Backpropagation (GA–BP) neural network with SHapley Additive exPlanations (SHAP), denoted GA–BP+SHAP. Here, backpropagation (BP) refers to the standard feedforward neural network trained via error backpropagation, while SHAP provides interpretable feature attributions. A domain-aware feature set includes Compliance Rating (CR) and Clinical Workflow Coupling (CWC), the latter operationalized via process mining on directly-follows graphs. The dataset comprises $N = 1{,}200$ projects (670 Medical Information Mart for Intensive Care (MIMIC)-derived, 530 industry). We enforce leakage-safe preprocessing, a time-aware split (last 20% as holdout test), and group-aware nested cross-validation (company as group) for model selection. To address tail imbalance, we use within-fold Synthetic Minority Over-sampling with Gaussian Noise (SMOGN) and a quantile-weighted mean square error (MSE). A genetic algorithm tunes hidden units, activation, and learning rate under a matched evaluation budget. Interpretability uses KernelSHAP with expert review. Results The GA–BP model outperforms strong baselines (Elastic-Net, Support Vector Regression (SVR), Random Forest, eXtreme Gradient Boost (XGBoost)/Light Gradient-Boosting Machine (LightGBM), plain BP), achieving test mean magnitude of relative error (MMRE) $\approx$ 21.9% and PRED(25) $\approx$ 83.3%, with gains significant after Holm–Bonferroni correction. Calibration/residual diagnostics show reduced bias, especially in the high-cost tail. SHAP highlights CR and CWC as dominant drivers. Expert assessment reports substantial agreement with model attributions (Cohen’s $\kappa \approx 0.73$, $p\;\lt\; 0.001$). Conclusions Combining domain-specific features with metaheuristic tuning and validated explainable artificial intelligence (AI) yields estimates that are both accurate and auditable. The framework supports budgeting, risk triage, and stakeholder communication in regulated settings. Limitations include concentration in three vendors and dependence on log quality for CWC; future work targets prospective, cross-organization validation and multi-objective optimization (effort, schedule, rework risk).