[abstract] YADEM3: A FUZZY SET MODEL OF PAIN-ONLY DECOMPRESSION SICKNESS IN SINGLE STAGE NO-D AIR DIVES - FIRST IMPRESSIONS

Abstract

BACKGROUND: Fuzzy logic is a formal system, although its name implies (incorrectly) little scientific rigor. Fuzzy logic is basically a superset of traditional (Boolean) logic, but allows partial true/false values. Thus, while Boolean logic allows logical values of 0 or 1 (FALSE or TRUE) only, fuzzy logic also allows values between 0 and 1. A pilot study was performed to examine the ability of fuzzy logic to model the occurrence of decompression sickness. METHODS AND MATERIALS: The study models the occurrence of decompression sickness as a function of two fuzzy variables depicting the depth of the dive (represented by 3 fuzzy sets: SHALLOW, MEDIUM and DEEP) and the bottom time of the dive (represented by 3 fuzzy sets: SHORT, MODERATE and LONG). The outcome is represented by the fuzzy sets: SAFE, QUESTIONABLE and UNSAFE. A 3 x 3 matrix is used to associate fuzzy depth and fuzzy bottom time by nine rules. The rules range from "IF DEPTH IS SHALLOW AND BOTTOM TIME IS SHORT, THEN THE DIVE IS SAFE" to "IF DEPTH IS DEEP AND BOTTOM TIME IS LONG, THEN, THE DIVE IS UNSAFE". The optimizing dataset consists of 907 single stage no-decompression stop man-dives on air to depths less than 200 fsw. RESULTS: The "best" fit model has a c index of 0.79, but with many ties. The sensitivity/specificity is 0.81/0.61 at an activation level of 0.13. All fuzzy sets use triangle membership functions. Trapezoid functions were less robust during optimization. CONCLUSIONS: Fuzzy set models of decompression sickness warrant further investigation in predicting DCS. Fuzzy logic holds the most promise when combined with another methodology to form a hybrid model. Examples would be the use of fuzzy set inputs to a neural network decompression model, or modeling a Haldanean model using fuzzy set tissue supersaturation ratios. Although the models developed in this study have some predictive ability, they cannot be used without validation with an independent data set.