Betrouwbare Machinaal Leren-Methoden voor de Analyse van Polysomnografieën

Machine Learning (ML) is widely applied to automate repetitive tasks, offering potential benefits across various domains. These include safety-sensitive applications such as the analysis of polysomnographies in sleep medicine. Although there are unmistakably many benefits of using machine learning for sleep analyses, concerns have been raised about the accuracy and robustness of the proposed models. This prevents the models from being integrated in the medical routines. We draw inspiration from concrete challenges posed by two applications in sleep analysis: sleep stage scoring and respiratory analysis. We propose multiple machine learning approaches for enhancing the trustworthiness and reliability of ML models. Moreover, we illustrate their efficacy on the two aforementioned medical tasks. We focus on three different aspects: the explainability of the model, the ambiguity of labels and predictions, and the robustness of the model. The primary barrier is the lack of explainability of the developed models that leads medical experts to be cautious about incorporating new learned models into their workflows. Because medical experts retain ultimate responsibility for care, even when models are accurate and time-saving, the lack of transparency (e.g., about what data was used to train the model, how it arrives at its prediction) still represents a significant hurdle. We address this concern in sleep stage scoring by proposing a classifier which mimics the way humans score which makes it simple to explain. Furthermore, we developed a novel method for multiple sequence alignment on sequences of item sets. This enables identifying pathological breathing patterns such as apnea and hypopnea without the need for complicated feature construction or classification methods. The second barrier is the ambiguity of the class labels in the data. Despite having access to a sufficiently large training set, an ML model can still make systematic mistakes. For some observations, it may, for example, not be possible to unambiguously associate a single class label such as for observations in a transition region between two classes. Moreover, the training data can contain too many incorrectly labeled examples. This ambiguity has two consequences which we tackle. First, there are examples for which the model is not able to predict a single class with high certainty. We resolve this by introducing a reject option into the sleep stage classifier, i.e., we allow the classifier to abstain from predicting a sleep stage if it is too uncertain in its prediction. Second, it is not possible to evaluate the model using standard techniques when the true class label for a test example is not unique. Hence, we developed an alternative performance evaluation that takes both the prediction and label uncertainty into account. The third barrier is the lack of robustness present in ML models when they are forced to extrapolate to regions of the feature space that were not (sufficiently) present in the training data. Such extrapolation is omnipresent in many practical cases in which a model is (unintentionally) used on data that deviates from the training set. In sleep medicine, for example, the data is affected by factors such as the patient's age, gender and pathologies. Hence, it is extremely difficult, if not impossible, to ensure that the training data is sufficiently diverse and exhaustive. We show that an anomaly detector is useful to identify deviating patients in sleep stage scoring. However, this approach results in a sub-optimal rejection performance when applied on individual predictions, i.e., the reject option is often too aggressive in practice as there are many seemingly anomalous patterns which are easy to classify accurately. This is, for example, the case when a patient is moving. In that case, multiple sensors register signals with very high and unique amplitudes while movement is often associated with wakefulness. Therefore, we introduce a new two-step rejection framework which includes a dependency between the reject option and the classifier itself to overcome this problem.