Multinomial Probabilistic Modeling for Decompression Sickness using Gas Content and Bubble Volume Models

Abstract

Decompression sickness (DCS) is a condition resulting from the decrease in ambient pressure, both in hyperbaric and hypobaric environments. When ambient pressure is decreased, inert gas dissolved in the body’s tissues can become supersaturation and form bubbles, the physiological precursor to DCS. The signs and symptoms of DCS range from mild joint pain or rash to serious cardiopulmonary or neurological dysfunction. DCS is treated with recompression therapy, in which a patient is recompressed in a hyperbaric chamber and then decompressed following a treatment schedule specific to their symptoms.This work focuses on DCS arising from hyperbaric exposures, and specifically underwater diving. DCS is a risk faced by U.S. Navy divers during underwater missions. To gain insight into the level of risk posed by a particular time-depth profile, the Navy uses probabilistic decompression models. These models quantify the probability of DCS occurring using survival analysis and a gas content or bubble volume model to define risk. Current probabilistic models make the binomial prediction of the probability of DCS occurring and not occurring.Risk is a two-faceted entity, in which both the probability of injury and severity of injury contribute to the level of risk posed by an activity. Many Department of Defense organizations use two-dimensional risk assessment matrices to consider both the probability and severity of injury to manage operational risk. However, the U.S. Navy’s probabilistic decompression models only predict the probability of injury and do not provide divers with any information regarding potential DCS symptom severity. The goal of this work was to investigate the efficacy of a variety of multinomial probabilistic models which simultaneously predict the probability and severity of DCS injury. We first conducted an analysis on the BIG292 calibration dataset to uncover the source of the bimodal trend in DCS symptom onset times. We concluded that neither of the two peaks alone correspond to DCS symptoms or DCS resulting from a particular dive type, and rather are a product of dive trial medical surveillance protocol. Because the dataset’s bimodal symptom onset time behavior is not related to the illness itself, it is not necessary for a probabilistic model to reproduce this trend.Next, we developed 20 multinomial probabilistic models, testing the effectiveness of gas content versus bubble volume models to calculate risk, the justification of various gas content model parameters, the impact of using Type A/B versus Type I/II symptom severity splitting methods to define mild/serious DCS cases, and the influence of treating marginal DCS cases as separate, hierarchical events versus considering them non-events during model calibration. The multinomial models presented in Chapters 3, 4, and 5 are able to accurately predict the incidence and severity of DCS as observed in the calibration data within 95% confidence. We find that trinomial models, which predict the probabilities of mild, serious, and no DCS, do perform better on the calibration dataset then their binomial counterparts. Multinomial models that predict the probabilities of marginal DCS are not able to accurately replicate the onset or distribution of marginal DCS cases in the calibration dataset, and the use of marginal cases during model calibration affects the model’s ability to accurately predict other types of DCS as well. The multinomial bubble volume models tested in this work were not able to achieve an optimal parameter set, and experienced model failure when predicting zero risk for a subset of dives. We do not recommend the further use of multinomial models that predict the probability of marginal DCS separately from other symptom types. The multinomial bubble volume models should be reoptimized with a different algorithm and/or an alternative bubble nucleation criterion. To determine which multinomial probabilistic model presented here is optimal for U.S. Navy dive planning, all models should be evaluated on data extraneous to the datasets used for model calibration.