Fig. 4: Evaluation metrics for studies classifying five sleep stages using wEEGs.

Each panel (a, c–g) includes a boxplot of metric distributions and a bar chart showing the top 20 studies by MCC score for that specific stage. Asterisks (*) indicate statistically significant differences (p < 0.05) from MCC (a, c–g) or between micro- and macro-averaged versions of the same metric in (b). a Overall metrics: Results demonstrated that overall accuracy (ACC) was significantly higher than Cohen’s kappa (κ) and MCC. b Comparison of micro- and macro-averaged metrics: Macro-averaged ACC was significantly higher than micro-averaged ACC, while micro-averaged SE, PPV, and F1 were significantly higher than their macro-averaged counterparts, highlighting class imbalance effects. c Wake stage: Wake showed the highest SP and NPV but the lowest SE among stages, indicating reliable exclusion of wake but variable detection. d N1 stage: N1 was the most challenging stage, with the lowest scores across κ, SE, PPV, F1, and MCC. e N2 stage: Showed relatively high metrics overall. Being the most prevalent class, the differences between metrics emphasizing TPs (e.g., SE) and TNs (e.g., SP) were smaller compared to other stages. Nevertheless, the MCC and κ scores were noticeably lower, suggesting potential misclassifications. f N3 stage: N3 showed the highest κ and MCC, with strong SE and excellent agreement across studies. g REM stage: REM had the second-highest SE and MCC but the lowest PPV, indicating a tendency for false positives. h Mean proportional distribution of sleep stages across studies: Wake (18%), N1 (7%), N2 (41%), N3 (17%), REM (17%). ACC: accuracy; κ: Cohen’s kappa; MCC: Matthews correlation coefficient; SE: sensitivity; SP: specificity; PPV: positive predictive value; NPV: negative predictive value.