A gradient-based optimization model for predicting decompression sickness risk.

Decompression sickness (DCS) is a low-incidence but potentially severe consequence of hyperbaric exposure. Probabilistic decompression models offer a framework to quantify this risk, yet their calibration is challenged by the scarcity of empirical outcome data. In this study, we propose a gradient-based optimization model to predict DCS probability, trained on 924 dive profiles from the US Navy Experimental Diving Unit XVal-He-9 tables, representing predefined DCS probabilities (2.3% and 4%), and optimized based on actual body tissues grouped in five compartments. The model achieved high predictive accuracy (MAE: 0.535%; RMSE: 0.694%) with consistent performance across training and test sets, indicating limited overfitting. Reduced accuracy was observed in intermediate depth ranges (100-130fsw or 30-39msw). Out-of-sample evaluation on 31 high-risk dives (three DCS cases) showed general agreement between predicted and observed incidence while suggesting a potential contribution of repetitive exposures not accounted for in the model. These results demonstrate that gradient-based optimization, trained based on existing probabilistic tables, seems to be capable of satisfactorily predicting decompression sickness risk for a given dive profile. Additionally, future studies can further adjust the loss function to account for individual or dive-related indicators, leading to a more individualized risk function.