Over the years, the ML algorithms have been improved to be more and more performant. In particular, deep learning methods advantageously capture complex patterns in data, therefore allowing to reach higher levels of accuracy (20). But concretely, clinicians expect more than just high accuracy from predictive systems, which opacity constitutes a constant criticism (16, 19). The emerging domain of explainable Artificial Intelligence (xAI) is of particular interest in this respect (21). Indeed, explainability allows clinicians to choose to trust, or not, the recommendations (22). Moreover, it is well established that ML models tend to reproduce the biases present in the training data sets, often caused by the unbalanced representation of the classification categories. Explainable decision chains thus allow to control that the outputs are conform to ethical standards, and notably unsupported by any form of discrimination (23, 24).

Most of the attempts in developing an xAI were focused on the design of post-hoc systems, i.e. black boxes completed by a component explaining predictions a posteriori (25). Post-hoc systems are thus thought as an interesting way of combining high accuracy and explainability. However, they remain questionable on (i) the veracity of their explanations, which are generated around the considered data point, and (ii) the consequent inability to give a comprehensive picture of the model behavior (21).

Concurrently, it has been shown that models, ranging from white to black boxes, all perform comparatively when trained on quality and meaningful data (21, 26). This observation suggests a double perspective on the development of explainable ML systems.

Efforts should thus be made both on data preprocessing and model design, in order to better address the need for explainability and transparency required by medical applications.

Two main methodologies exist for scientific modeling (29).

While theory-based methods are usually considered for the understanding of disorders, data-driven methods are rather considered for the design of clinical tools (16). A hybrid approach guided by data and theory would broaden the field of investigation, reconciling the existing scientific knowledge with elements extracted from data. The concept, which is not new, was highlighted in (16), and then properly formalized in (29) as the Theory-Guided Data Science (TGDS) paradigm. This principle should be encouraged in psychiatric research. Indeed, TGDS may be put in practice through the interaction with domain experts (i.e. psychiatrists, neuroscientists, neurologists) bringing their medical knowledge for feature selection (16), or more globally to refine ML models in the frame of an expert-in-the-loop mechanism (30). The aforementioned explainability naturally fosters the implementation of a TGDS.

Though they are mostly considered in the development of decision aid systems, Multi-Class Classification (MCC) algorithms are criticized for several reasons. Indeed, MCC does not address comorbidity appropriately since it considers the different diagnostic categories as mutually exclusive (16). In addition, MCC becomes more challenging in presence of unbalanced and noisy data sets (31). The domain of One-Class Classification (OCC) (32) covers a range of algorithms capable of describing a given class [e.g., a neuropathology (33, 34)], in such a way to reject cases that do not comply with this description. It is thus possible to use ensembles of one-class models in order to test a patient for several conditions simultaneously. One-class classification also gets rid of the need for a balanced data set including training instances from each class, as required by the MCC scheme. Finally, through its very nature, OCC can efficiently rule out noise, and specifically class noise which is usually located on the class boundaries (31, 35).

Several OCC tools are already available for clinical use, despite being mainly targeted towards neurological disorders (36). Hence, OCC deserves greater attention to be further developed in psychiatric research. Additional efforts for algorithmic improvements would be particularly worth considering in the context of explainable AI.

Though outstanding, the efforts for large-scale data gathering across several sites yield disparities in terms of demographics and experimental protocols (22). The homogenization of experimental protocols, and the design of appropriate validation procedures are respectively thought as ways for achieving and assessing generalizability (18, 19). The question of the extent to which this generalizability needs to be achieved deserves to be discussed, and probably requires a scientific consensus to provide a clear research direction. Indeed, the available financial and technical means for medical assessment and data may differ from a region to another.

Furthermore, psychiatric conditions can be characterized by clinical and/or neurobiological heterogeneity, the latter being also established in healthy controls (37). In this case, a thorough analysis may help to consider the best modeling strategy. For example, a ML framework can be implemented to perform diagnosis prediction in different levels, i.e., to detect a disorder first, and the disorder subtype then (38, 39). In (33), the authors focused on the description of ASD through OCC, since controls showed high neurobiological disparity. Moreover, ahead of the ML process, the experimental protocol should not necessarily be aligned with the DSM diagnostic categories. Indeed, these diagnostic labels are heterogeneous and derived from traditional assessments conducted by clinicians (40, 41). The Research Domain Criteria (RDoC) framework was introduced by the National Institute of Mental Health to alleviate this issue (42, 43). The RDoC orients the study of mental illnesses towards domains of human functioning described at different levels, rather than towards symptoms. The lowest level relates to units of analysis, suggesting relevant biological, genetic and physiological investigation markers (43).

The psychiatric domain has witnessed a significant increase in ML-based studies, along with a diversification of the modalities considered for data processing and modeling (44). It is therefore imperative to apply guidelines that ensure reproducibility; recommendations are provided in (44). The appropriate choice of procedures for model training and evaluation, as well as the availability of source code/data should notably be encouraged. Yet, a recent study highlighted that these aspects are often lacking: it appeared that 50% of studies do not share software, while 36% do not give access to data (45).

More specifically in the context of open data sharing, a standard segmentation of the data collections into training and test sets would reinforce reproducibility (40). On the occasion of the ADHD-200 competition, training and test sets were kept separately in order to allow respectively the development and the assessment of the predictive models developed by the competing teams. Since then, these data subsets have mostly been used in their initial form, which makes it easy to report the evolution of the progress achieved on the prediction of ADHD. The same cannot be said for other INDI data sets such as the ABIDE collection, where the segmentation of data is a choice made for each research study. This great disparity in the definition of the data subsets therefore makes it difficult to track the research progress on a given mental disorder.