Hybrid deep learning and statistical modeling for interpretable disease progression prediction in medical sciences

One of the significant problems in medical informatics is the correct prediction of the course of a disease because of the nonlinear and irregular nature of patient course and time dependence. Current models like Random Forest, SVM, CNN-LSTM and Copula-Based Random Forest have a plausible predictive strength. Nonetheless, either they depend on static features or cannot be clinically interpreted. The proposed research study presents the concept of a hybrid framework AdaGoDE that combines a Latent Ordinary Differential Equation (Latent-ODE) encoder with an Adaptive Generalized Additive Model (Adaptive-GAM) decoder. The objective is to formulate disease progression in a continuous time environment and still have predictions that are interpretable. The first one is to produce patient-specific latent trajectories and decode them into lexical clinical outcomes with the help of adaptive spline functions. The innovation is that continuous-time latent dynamics are being combined with adaptive penalized smoothing to allow strong modeling of continuous-time dynamics as well as interpretations that are meaningful in clinical practice. The methodology consists of the encoding of patient information as the latent trajectories, using the Latent-ODE encoder, and the decoding of progression scores, using the Adaptive-GAM with patient-specific smoothing parameters. The implementation was done in Python and TensorFlow libraries. The Parkinson Disease Voice Dataset with 42 patients and 5875 longitudinal voice recordings was used as the evaluation tool. Key performance metrics included MAE, RMSE, R 2 , Concordance Index, and Temporal Smoothing Index. Results demonstrated a mean MAE of 0.030, RMSE of 0.045, and R 2 of 0.985, surpassing baseline models while maintaining interpretability. The results show that AdaGoDE makes precise, seamless, and clinically explainable predictions, which proves its opportunities as a valid framework of personalized disease tracking and progressive prediction in Parkinson disease. • Latent-ODE encoder captures continuous-time disease dynamics for high-resolution, patient-specific trajectories. • Adaptive-GAM decoder links latent states to clinically meaningful markers, improving transparency and interpretability. • AdaGoDE outperforms baselines with MAE 0.030 and R² 0.985, delivering accurate progression prediction.