An Explainable Hybrid Deep Learning System for Medical Insurance Premium Optimisation to Minimise Out-of-Pocket Shortfalls in Zimbabwe

The context regarding the issue of medical insurance coverage in Zimbabwe can be considered through the perspective of critical coverage rates (10–15%) and out-of-pocket (OOP) expenses contributing to 37% of total healthcare expenditure, along with actuarial sustainability problems as a result of economic instability, hyperinflation, and fluctuations in the exchange rate. Conventional actuarial methodologies applied to assess risk factor interaction are unable to tackle such a problem efficiently. Thus, the Explainable Artificial Intelligence Premium Optimisation System (XAI-POS) based on the implementation of LSTM, CNN algorithms, and RL for modelling health risks and premium optimisation is suggested as a solution for the research problem. The dataset included 1,500 patient records with 87 variables from 2020 to 2025 (IPEC Administrative Returns, Ethics Clearance: HIT-REC-2024-0147; IPEC Data Access Agreement: IPEC-DAA-2024-009). Temporal train/test split approach for avoiding data leakage in the cross-sectionally divided dataset was utilised since the latter could potentially affect the results obtained (the 2020-2022 dataset with n=900 patients became the training set, while 2023-2025 dataset with n=600 patients – test set). The RL agent underwent training via 500 episodes using the epsilon-greedy algorithm. Some of the major results obtained include: 0.88 ± 0.02 for an individual healthcare cost prediction model Rsquared on the testing dataset, a decrease by 18.9 percent in average OOP gap, decrease of 11.5 percentage points in the number of catastrophic expenditure cases, affordability rating for premium of 74 out of 100, actuarial loss ratio between 0.60 and 0.85 for sustainable level, and decrease in premium Gini coefficient from 0.412 to 0.331.