Interpretable Machine Learning Assessment

Mislabeled learning for psychiatric disorder detection

“…Although deep neural networks (DNN) and ensemble learning methods such as random forests (RF) and extremely randomized trees (Extree), can achieve decent performance also, they especially lack good reproducibility for their built-in randomness, which is essential for clinical psychiatric diagnosis [56][57][58][59]. We further employ nonnegative singular value approximation (nSVA) for SNP feature selection for its proven effectiveness and efficiency for high-dimensional data [60][61].…”

“…Since the traditional classification measures are neither efficient nor interpretable in assessing different machine learning models' performance, we extend the proposed diagnostic index (d-index) measure under binary classification to provide a more explainable and sensitive learning performance evaluation [39][40]. This is because the traditional classification measure assessment may only reflect one aspect of classification performance.…”

“…The d-index definition of the binary classification can be found in the following section and more d-index information can be found in [40]. ) if there is no underfitting [40]. [15,30,31].…”

“…However, it would potentially create a learning risk because it is likely that some subtypes have only a few samples. The scarcity of samples in a certain type will generate inaccurate or even biased learning results under the situation [25,40].…”

“…The maximum value of d-index is 2 where classification has the perfect results. The minimum value of the d-index is 2𝑙𝑜𝑔 5 ( F 5 ) if there is no underfitting [40].…”