Research Reveals Targeted, High-Accuracy AI Outperforms Complex Models In Mental Health Tech

The effectiveness of AI cannot always be attributed to greater complexity. Sometimes, the answer lies in selecting techniques tailored to a task’s specific requirements. Such an approach could potentially yield more reliable and explainable results than a complex model. A case in point is suicide risk assessment.

A team of researchers comprising Ahmedabad University alumni Shlok Hiren Shelat and Shrey Salvi, and Professor Souvik Roy, suggests that this approach could contribute to more transparent AI systems for assessing suicide risk from social media text. In this setting, models need to do more than predict risk severity. They must also identify the evidence supporting a prediction and recognise clinically relevant risk and protective factors. This is particularly challenging when available data is limited, and the consequences of an incorrect assessment are significant.

Their work, titled Bag of Tricks or Bag of Myths? Reducing Modelling Complexity with Task Knowledge in Explainable Suicide Risk Assessment was conducted in collaboration with Manas Gaur of the University of Maryland and under the mentorship of Professor Amit Sheth, Founding Director of the Indian AI Research Organisation (IAIRO). It addresses a common assumption in machine learning: that adding techniques such as larger models, synthetic data, loss reweighting, model ensembling, or threshold tuning will necessarily improve performance.

The team tested this assumption using 1,635 clinician-annotated social media posts. They evaluated 31 pre-specified techniques across seven methodological families through approximately 300 controlled experiments, using author-disjoint data partitions to reduce the risk of models learning patterns associated with individual authors. The results showed that these techniques do not necessarily improve performance simply because they are commonly used. Only five of the 31 comparisons produced reliable gains.

Based on these findings, the team developed a system around task-conditioned technique selection. The approach retains a technique only when it is suited to the task. This is determined by the nature of the problem, task-specific knowledge, and experimental evidence.

The system was designed to produce three outputs: a four-level suicide risk assessment, evidence supporting the assessment, and 24 clinical risk and protective factors. The team combined machine learning with task-specific rules and improved the system’s calibration, making its predictions more reliable. The final system achieved a composite score of 0.7781, with strong performance across risk assessment, evidence identification, and detection of clinical risk and protective factors.

For this work, the team received first position in the Explainable Suicide Risk Detection Challenge at the IEEE International Conference on Big Data 2026. The team’s broader effort, Neurosymbolic Intelligent Diagnostics for Holistic Trauma-Informed Care (NIDHI), ranked first overall among 53 teams.

The work reveals how domain knowledge and systematic experimentation can guide the design of AI systems in settings where data is limited, and errors can have significant consequences. Rather than adding complexity by default, the research focuses on identifying which techniques are justified for a particular task and how different components should work together.