A gradient-based optimization model for predicting decompression sickness risk

Figures

• 

  Figure 1. Total loss training trajectory logged every 100 epochs.

• 

  Figure 2. Predicted P(DCS) according to bottom time and depth for the dive profiles with 4% observed P(DCS). A lower accuracy is observed in the 100fsw (30msw) to 130fsw (39msw) profiles, in an exacerbation of the effect observed in the lower P(DCS) profiles.

• 

  Figure 3. Error distribution for each expected P(DCS) group. The accuracy of the overall model is not the same for different expected P(DCS), probably due to the lower accuracy observed in the 100fsw (30msw) to 130fsw (39msw) range, which was more pronounced in the 4% P(DCS) dataset, as can be viewed in Figure 5. Additionally, shorter bottom times tend to produce slightly lower accuracy in the group with higher expected P(DCS), as demonstrated in Figure 4, where P(DCS) tends to be slightly overestimated. This performance difference can be observed when errors for different training sets are calculated separately (Table 5). It is clear that the lack of performance in the range above [100fsw (30msw) to 130fsw (39msw)] in the 4% P(DCS) training set affects the overall prediction accuracy for P^j.

• 

  Figure 4. Error as a function of bottom time according to the overall length of the dive. Shorter bottom times tend to result in a slightly lower accuracy for the higher estimated P(DCS) group of profiles. Each colored line is a LOESS-smoothed trend curve of the absolute prediction error as a function of bottom time.

• 

  Figure 5. Error versus depth for each group of expected P(DCS) as calculated for the test set. As can be observed, the outliers are concentrated in the range of 100fsw to 130fsw, for both expected P(DCS). Each colored line is a LOESS-smoothed trend curve of the absolute prediction error as a function of bottom time.

• 

  Figure 6. Sum of compartment inert gas pressures for all dive profiles used in the out-of-sample test. The red lines represent the dive profiles with higher expected P(DCS).

• 

  Figure 7. Total compartment pressure for the 31 dives analyzed. Orange lines are used to highlight the moments when the linear gas washout was triggered during decompression. Linear kinetics is only observed in compartment 2 since PXO was defined as zero for compartments 1, 3, 4, and 5.

• 

  Figure 8. Bootstrap estimation of P(DCS) for the 31 profiles.

• 

  Figure 9. Dives that resulted in DCS and dives made in the previous 24- to 48-h intervals by the same divers.