Knowledge-Infused Learning for Developing a Mental Health Diagnostic Copilot in Healthcare Systems

Gerald Ketu NdawulaAAAI 2025 · DOI

Introduction

In the current healthcare landscape, mental health diagnostics face a dual challenge: a critical shortage of specialized providers and a lack of adaptive, culturally competent diagnostic tools. While Artificial Intelligence (AI) has been integrated into various medical fields, existing solutions often function as "black boxes" or static recommenders. These systems frequently fail to capture the nuanced, evolving nature of psychiatric conditions and often overlook the specific needs of underserved populations.

Presented at AAAI-25, this research proposes a "diagnostic copilot" designed to move beyond static recommendations. By combining Knowledge-Infused Learning (KIL) with multi-turn conversational AI, the author outlines a framework for a system that is clinically grounded, adaptable to real-time patient feedback, and specifically engineered to reduce diagnostic disparities in marginalized communities.

The Limitations of Current Healthcare AI

The paper identifies a fundamental "static" problem in contemporary healthcare AI. Models like IBM’s Watson or Google’s Med-PaLM typically provide one-off recommendations based on historical data. However, mental health diagnosis is rarely a single event; it requires ongoing reassessment and interactive dialogue.

Furthermore, a significant "equity gap" persists. For example, African Americans are 20% more likely to experience serious mental health issues but are significantly less likely to receive adequate care. When AI models are trained on biased datasets without clinical grounding, they risk perpetuating these disparities. The author argues that a diagnostic tool must be a collaborative partner—a copilot—that engages both patients and clinicians in a continuous feedback loop.

Methodology: Knowledge-Infused Learning (KIL)

The core of the proposed system is the integration of structured medical knowledge into the machine learning pipeline. This approach, known as Knowledge-Infused Learning (KIL), ensures the AI is not merely predicting the next word in a sentence but is operating within the bounds of established medical science.

1. Architectural Integration

Instead of treating medical knowledge as an external lookup table, the system extracts structured data from trusted ontologies such as SNOMED CT and UMLS. This information is used to construct a comprehensive knowledge graph linking symptoms to potential diagnoses. The author suggests modifying the internal logic of the models—such as the splitting criteria in decision trees and ensemble models—to align with these established medical hierarchies.

2. Multi-turn Conversational Feedback

To mimic the diagnostic process of a human clinician, the copilot utilizes Natural Language Processing (NLP) to engage in interactive loops. Rather than accepting an initial patient statement as final, the system asks clarifying follow-up questions to refine its suggestions in real-time. This grounding in medical ontologies helps prevent the "hallucination" problem common in standard Large Language Models (LLMs).

3. Clinician-in-the-Loop (Instructability)

The system is designed to be "instruct-able," allowing healthcare professionals to guide its learning. By employing Direct Preference Optimization (DPO) or Kahneman-Tversky Optimization (KTO), clinicians can provide direct feedback on the model's suggestions. The AI then uses this feedback to improve its diagnostic accuracy through reinforcement learning, ensuring that the final output aligns with clinical best practices.

Data and Training

To ensure the model understands both clinical terminology and the natural language used by patients, the copilot is designed to be fine-tuned on a diverse set of specialized datasets:

  • ESCON: Emotional Support Conversation dataset.
  • IHMI: Interpretable Mental Health Instruction dataset.
  • SMHD: Self-reported Mental Health Diagnoses.
  • Primate: Process Knowledge for Depression Triage.

This mix of electronic health records and real-world patient-provider conversations allows the model to bridge the gap between formal medical knowledge and the lived experience of the patient.

Evaluation Framework

As a research proposal, the paper focuses on a robust evaluation framework to measure the copilot's efficacy across several dimensions:

  • Clinical Entity Overlap Score: This measures how accurately the AI identifies clinically relevant concepts compared to a "gold-standard" knowledge graph.
  • Logical Coherence: Utilizing RoBERTa-based Natural Language Inference (NLI), the framework checks that the AI’s conversational outputs do not contradict themselves or established medical facts.
  • Semantic Relatedness: BERTScore is used to ensure that the AI's responses are contextually accurate and relevant to the patient's specific input.
  • Instructability: The framework quantifies how effectively the model adapts to clinician corrections via DPO/KTO, measuring the speed and accuracy of its improvement.

Conclusion and Future Directions

The proposed diagnostic copilot represents a shift from AI as a simple recommender to AI as a collaborative partner. By grounding generative models in medical ontologies, the author addresses the critical challenges of hallucinations and data bias, particularly in "low-data" scenarios involving marginalized populations.

While the integration of diverse data sources (ontologies, EHRs, and live chat) presents technical challenges regarding latency and complexity, the potential for "traceable reasoning" offers a path toward greater trust between patients, providers, and AI systems. Future work will focus on scaling this model to other clinical domains and conducting extensive user studies to validate its impact on diagnostic equity and accuracy.

Summary updated

BibTeX
@inproceedings{ndawula2025knowledge,
  author = {Gerald Ketu Ndawula},
  title = {{Knowledge-Infused Learning for Developing a Mental Health Diagnostic Copilot in Healthcare Systems}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i28.35335},
}
Figure