DESIGN AND SIMULATION-BASED EVALUATION OF AN INTERPRETABLE FUZZY LOGIC FRAMEWORK FOR SLEEP APNEA SEVERITY CLASSIFICATION
Abstract
Obstructive sleep apnea (OSA) is graded by the apnea–hypopnea index (AHI), yet the grade a patient receives depends on a handful of hard cut-offs that sit awkwardly on top of noisy, partly overlapping physiological signals. Machine-learning classifiers deal with that noise but rarely show their reasoning, which makes them difficult to defend in a clinical setting. We designed a Mamdani fuzzy inference system (FIS) that grades OSA severity from four overnight oximetry and arousal measures, and we asked how much accuracy such a transparent system gives up, if any, relative to crisp threshold rules and standard machine-learning models. A synthetic cohort of 1,000 adults was generated from an explicit data-generating model in which the oxygen desaturation index (ODI), SpO2 nadir, time below 90% saturation (T90) and arousal index depend on a latent AHI with realistic measurement noise. An expert-specified FIS (28 rules, trapezoidal membership functions) was refined by differential evolution on a 70% training split. Comparators were three threshold classifiers and five machine-learning models. Evaluation used a 30% hold-out set (n = 300) with bootstrap confidence intervals and exact McNemar tests, five-fold cross-validation, feature ablation, and membership-function sensitivity analysis. On the hold-out set the expert FIS reached 70.3% accuracy (95% CI 65.0–75.3) and the refined FIS 80.7% (76.0–85.0), against 74.7% for ODI thresholds at the conventional AHI cut-offs and 81.3–83.7% for logistic regression, k-nearest neighbors, SVM and random forest. The refined FIS did not differ significantly from any of these four models (McNemar p ≥ 0.24) and made no error larger than one severity grade. Under five-fold cross-validation the machine-learning models stayed four percentage points ahead (82.2–83.9% versus 79.3%). Accuracy was sensitive to the position of the SpO2 membership functions and comparatively insensitive to those of T90 and arousal index. In this simulation, a refined fuzzy system recovers most of the accuracy of black-box models while exposing every rule that contributed to a decision. Because the data are synthetic and ODI was generated as a close proxy of AHI, the findings describe the behaviors of the method and should not be read as clinical performance; validation on polysomnography recordings is the necessary next step.
Keywords
- Sleep apnea; fuzzy logic; Mamdani inference; interpretable machine learning; severity classification; pulse oximetry; simulation study.
Views & Downloads Statistics (Last 12 Months)
Monthly performance and trends of article views and downloads
