Optimising Pharmaceutical Formulations in Resource‑Limited Settings: A Hybrid Deep Learning Approach for Sustained‑Release Tablets

Introduction: Manufacturers in LMICs meet less than 20% of the demand for medical devices. Formulations of sustained-release tablets by means of design of experiments require 50-100 (costly) trials ($10k-15k per trial). Existing machine learning approaches (FormulationNet and PharmaML) depend on large proprietary datasets and GPU training. We intended to develop a CPU-based approach capable of predicting excipient ratios with a macro-average R² ≥ 0.90.

Methods: A literature review yielded 151 formulations. Training was augmented with Gaussian perturbation and linear interpolation to 5,440 training samples. The test set (20%) was not augmented. A residual attention CNN was used to predict the ratios of polymer, binder, diluent, lubricant, and glidant. The ratio components with single-model R² below 0.85 were supplemented with a shallow ensemble MLP (three models, mean blend). A Streamlit app (DeepPharm®) was used for the CPU-only inference.

Results: CNN provided R² = 0.933 and RMSE = 0.049. The ensemble boost increased the macro-average R² to 0.941 (RMSE 0.044), with component-specific improvements for binder_ratio (0.812 to 0.877), lubricant_ratio (0.803 to 0.874), and glidant_ratio (0.798 to 0.871). SHAP analysis revealed that the fraction of polymer (|SHAP| = 0.42), kinetic R² (0.38), and API logP (0.31) were the three main predictors. The average latency of the end-to-end model on an Intel i5-10400 CPU was 420 ms (SD 35 ms).

Discussion: Based on the results, we expect a potential saving of ≥80% of experimental iterations and development timelines, from months to weeks (prospective validation is pending).