Performance and utility trade-off in interpretable sleep staging

Abstract

Recent advances in deep learning have led to the development of models approaching the human level of accuracy. However, healthcare remains an area lacking in widespread adoption. The safety-critical nature of healthcare results in a natural reticence to put these black-box deep learning models into practice. This paper explores interpretable methods for a clinical decision support system called sleep staging, an essential step in diagnosing sleep disorders. Clinical sleep staging is an arduous process requiring manual annotation for each 30s of sleep using physiological signals such as electroencephalogram (EEG). Recent work has shown that sleep staging using simple models and an exhaustive set of features can perform nearly as well as deep learning approaches but only for some specific datasets. Moreover, the utility of those features from a clinical standpoint is ambiguous. On the other hand, the proposed framework, NormIntSleep demonstrates exceptional performance across different datasets by representing deep learning embeddings using normalized features. NormIntSleep performs 4.5% better than the exhaustive feature-based approach and 1.5% better than other representation learning approaches. An empirical comparison between the utility of the interpretations of these models highlights the improved alignment with clinical expectations when performance is traded-off slightly. NormIntSleep paired with a clinically meaningful set of features can best balance this trade-off by providing reliable, clinically relevant interpretation with robust performance.

An Interpretable Machine Learning Framework for Rare Disease: A Case Study to Stratify Infection Risk in Pediatric Leukemia

“…For this reason, the accuracies of the predicted infection risk may seem low compared to predictive models for other disease domains. For example, other clinical domains like sleep staging [ 28 ] or epilepsy [ 59 ] have shown extremely high accuracy with similar methodologies. However, this is because such models employ much larger sample sizes with many more granular features than were included in the present study’s rare pediatric leukemia dataset.…”

“…Interpretable and/or explainable methods make it easier to see why the model is making a particular prediction. It is possible that less interpretable black box methods might make better predictions [ 28 ]. However, black box methods that employ large neural networks need very large sample sizes, often more than 10,000 patients [ 59 , 78 ].…”

“…However, black box methods that employ large neural networks need very large sample sizes, often more than 10,000 patients [ 59 , 78 ]. There is often a trade-off between accuracy and interpretability [ 28 ], but advances in ML are narrowing this gap [ 59 , 78 ]. Use of probabilistic generative models—Another possibility for future work to apply ML to small tabular datasets for rare disease is probabilistic generative models.…”

“…However, black box methods that employ large neural networks need very large sample sizes, often more than 10,000 patients [ 59 , 78 ]. There is often a trade-off between accuracy and interpretability [ 28 ], but advances in ML are narrowing this gap [ 59 , 78 ].…”

“…As such, there is a known trade-off between less interpretable black box methods, which may be more accurate, and more interpretable glass box methods, which may result in slightly lower performance accuracy. Here, “interpretability” is defined by the transparency of the underlying ML model decisions and especially how the model’s decisions can be explained by real-world domain expertise [ 28 ].…”