What can we learn about the psychiatric diagnostic categories by analysing patients' lived experiences with Machine-Learning?

BACKGROUND: To deliver appropriate mental healthcare interventions and support, it is imperative to be able to distinguish one person from the other. The current classification of mental illness (e.g., DSM) is unable to do that well, indicating the problem of diagnostic heterogeneity between disorders (i.e., the disorder categories have many common symptoms). As a result, the same person might be diagnosed with two different disorders by two independent clinicians. We argue that this problem might have resulted because these disorders were created by a group of humans (APA taskforce members) who relied on more intuition and consensus than data. Literature suggests that human-led decisions are prone to biases, group-thinking, and other factors (such as financial conflict of interest) that can enormously influence creating diagnostic and treatment guidelines. Therefore, in this study, we inquire that if we prevent such human intervention (and thereby their associated biases) and use Artificial Intelligence (A.I.) to form those disorder structures from the data (patient-reported symptoms) directly, then can we come up with homogenous clusters or categories (representing disorders/syndromes: a group of co-occurring symptoms) that are adequately distinguishable from each other for them to be clinically useful. Additionally, we inquired how these A.I.-created categories differ (or are similar) from human-created categories. Finally, to the best of our knowledge, this is the first study, that demonstrated how to use narrative qualitative data from patients with psychopathology and group their experiences using an A.I. Therefore, the current study also attempts to serve as a proof-of-concept. METHOD: We used secondary data scraped from online communities and consisting of 10,933 patients’ narratives about their lived experiences. These patients were diagnosed with one or more DSM diagnoses for mental illness. Using Natural Language Processing techniques, we converted the text data into a numeric form. We then used an Unsupervised Machine Learning algorithm called K-Means Clustering to group/cluster the symptoms. RESULTS: Using the data mining approach, the A.I. found four categories/clusters formed from the data. We presented ten symptoms or experiences under each cluster to demonstrate the practicality of application and understanding. We also identified the transdiagnostic factors and symptoms that were unique to each of these four clusters. We explored the extent of similarities between these clusters and studied the difference in data density in them. Finally, we reported the silhouette score of + 0.046, indicating that the clusters are poorly distinguishable from each other (i.e., they have high overlapping symptoms). DISCUSSION: We infer that whether humans attempt to categorise mental illnesses or an A.I., the result is that the categories of mental disorders will not be unique enough to be able to distinguish one service seeker from another. Therefore, the categorical approach of diagnosing mental disorders can be argued to fall short of its purpose. We need to search for a classification system beyond the categorical approaches even if there are secondary merits (such as ease of communication and black-and-white (binary) decision making). However, using our A.I. based data mining approach had several meritorious findings. For example, we found that some symptoms are more exclusive or unique to one cluster. In contrast, others are shared by most other clusters (i.e., identification of transdiagnostic experiences). Such differences are interesting objects of inquiry for future studies. For example, in clear contrast to the traditional diagnostic systems, while some experiences, such as auditory hallucinations, are present in all four clusters, others, such as trouble with eating, are exclusive to one cluster (representing a syndrome: a group of co-occurring symptoms). We argue that trans-diagnostic conditions (e.g., auditory hallucinations) might be prime targets for symptom-level interventions. For syndrome-level grouping and intervention, however, we argue that exclusive symptoms are the main targets. CONCLUSION: Categorical approach to mental disorders is not a way forward because the categories are not unique enough and have several shared symptoms. We argue that the same symptoms can be present in more than one syndrome, although dimensionally different. However, we need additional studies to test this hypothesis. Future directions and implications were discussed. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1186/s12888-022-03984-2.