Predicting the future of neuroimaging predictive models in mental health

Abstract

Predictive modeling using neuroimaging data has the potential to improve our understanding of the neurobiology underlying psychiatric disorders and putatively information interventions. Accordingly, there is a plethora of literature reviewing published studies, the mathematics underlying machine learning, and the best practices for using these approaches. As our knowledge of mental health and machine learning continue to evolve, we instead aim to look forward and "predict" topics that we believe will be important in current and future studies. Some of the most discussed topics in machine learning, such as bias and fairness, the handling of dirty data, and interpretable models, may be less familiar to the broader community using neuroimaging-based predictive modeling in psychiatry. In a similar vein, transdiagnostic research and targeting brain-based features for psychiatric intervention are modern topics in psychiatry that predictive models are well-suited to tackle. In this work, we target an audience who is a researcher familiar with the fundamental procedures of machine learning and who wishes to increase their knowledge of ongoing topics in the field. We aim to accelerate the utility and applications of neuroimaging-based predictive models for psychiatric research by highlighting and considering these topics. Furthermore, though not a focus, these ideas generalize to neuroimaging-based predictive modeling in other clinical neurosciences and predictive modeling with different data types (e.g., digital health data).

Fig. 1. The increasing difficulty of understanding biases as application complexity increases.

In theoretical work, such as algorithmic proofs, bias is low, putatively, as these works often do not focus on real-world data. However, biases quickly emerge in well-established applications in machine learning, like language and image processing. These biases may be missed during the initial product development but can quickly become apparent upon widespread use. Finally, for emerging applications of machine learning, such as in psychiatry, potential biases are often hard to observe, understand, and prevent, in part because (1) our knowledge of mental health disorders is still limited in comparison to traditional applications like image processing and (2) the data may not be comprehensive enough to model the complexities of mental health fully.

Fig. 2. Benefits and trade-offs in using different models.

While exceptions may exist, the interpretability of models usually occurs at the price of prediction performance and vice versa. In neuroscience research, including understanding the neural circuits underlying psychiatric disorders, interpretability (defined as the ability to understand the cognitive and neurobiological underpinnings of a model’s features) offers the greatest utility. In contrast, prediction performance is a priority in many real-world products and applications. We argue that the target goal for neuroimaging-based predictive models on the interpretability/prediction performance trade-off is on the interpretability side. At the moment, interpretable predictive models are expected to better advance neuroimaging research in psychiatry and complement traditional approaches than models that sacrifice interpretation for prediction performance.

Fig. 3. Transdiagnostic prediction.

Current theories postulate that symptoms lie on a continuum, where distinct symptoms group together in overlapping clusters. As a result, and as discussed in “Dirty data”, real-world patients often exhibit many different patterns of symptoms and comorbidities rather than a single distinct pattern. Such viewpoints make classification into textbook diagnoses difficult as these diagnoses are based on meeting exemplar symptom patterns. Predictive models offer a solution to transdiagnostic problems, either by placing an individual into a cluster of patients that most mimic their spectrum of symptoms (i.e., transdiagnostic clustering) or by identifying brain networks that predict symptoms and generalize across a spectrum of traditional clinical categories and “healthy” individuals (i.e., transdiagnostic regression).