Oximetry Risk

Analysis of nocturnal paediatric oximetry uses trained Support Vector Machines (SVMs). If Regression is selected, a Support Vector Regressor (SVR) is used to predict the AHI. If Classification is selected, a Support Vector Classifier (SVC) assigns a risk classification.

Regression

Predicted AHI

If Regression is selected, an SVR produces a point estimate of the apnoea-hypopnea index (AHI). Measures of uncertainty are provided through the reporting of confidence intervals calculated using a quantile regression model fit using the residuals in testing of holdout Queensland Children's Hospital (QCH) and external PATS polysomnography data.

Classification

Distance

If Classification is selected, an SVC calculates the unitless distance of the final prediction to the decision boundary (scale: -inf to +inf, with 0 representing the boundary). The signedness is matched to the final prediction: negative values represent a negative result, and positive values a positive result. The magnitude of the distance represents the degree of belonging (to the assigned class) but is not a probability.

Calibrated Probability

The reporting of calibrated probabilities allows clinicians 1) to relate the Distance value to the probability of the real-world outcome, and 2) to understand the uncertainty surrounding the prediction using confidence intervals. The calibrated probability of an AHI of ≥5 is calculated by comparing the probability of the real occurrences of the event against the Distance score provided by the SVC, through testing of the SVC on real world data; a combination of holdout QCH and external PATS polysomnography data.

Explanations

Feature importance values or explanations attempt to describe the impact of the feature on the classification decision.

Global Explanations

Shapley value approximations are used to estimate the overall importance of each feature to the classification. They are approximated using Monte Carlo resampling of the model support vectors and are the mean of the marginal contributions of each feature. The size of the absolute value is a measure of the overall importance of that feature, and the signedness of the impact value indicates the directionality of the relationship to the prediction.

Local Explanations

Calibrated probabilities allow clinicians to 1) relate the Distance value to the probability of a real-world outcome, and 2) understand the uncertainty surrounding the prediction through confidence intervals. The calibrated probability of an AHI of ≥5 is calculated by comparing the actual occurrence of the event with the Distance score computed by the SVC. Calibration of the SVC was performed on holdout QCH data and external PATS polysomnography data.