Identification of Neurodevelopmental & Mental Health Conditions: An AI Approach to Reducing Barriers

This project builds on previous work that developed a tool to guide users through the diagnostic criteria for autism. It utilized Natural Language Processing (NLP) APIs to process open-ended responses.

Autism, ADHD, OCD, and GAD

Autism is a lifelong neurodevelopmental disability that frequently co-occurs with other neurodevelopmental and mental health conditions, including attention deficit hyperactivity disorder (ADHD), obsessive-compulsive disorder (OCD), and generalized anxiety disorder (GAD). These comorbidities complicate diagnosis and daily functioning, as symptoms may overlap, conflict, or amplify autistic traits.

Barriers to Diagnosis

Many encounter significant barriers to formal assessment, including long waitlists, high costs, and clinician bias. Additionally, clinical assessments often prioritize behavioural observation over structured self-report. These barriers contribute to underdiagnosis in adults, BIPOC individuals, and women.

Self-Diagnosis & Current Resources

When formal assessment is inaccessible, many turn to self-diagnosis. While helpful for self-understanding, challenges like imposter syndrome and difficulties obtaining accommodations are common.

Existing resources rely on wordy, single-disorder questionnaires and fixed Likert-scale questions, limiting their ability to capture nuances.

Purpose

This project developed an AI-assisted tool to evaluate whether a user’s experiences align with diagnostic traits associated with these disorders. It serves as a complementary or alternative resource to existing tools like the RAADS-R.

Engineering Goals

Promote Self-Reflection: Help users determine if a formal assessment may be beneficial.

Enhance Self-Report Reliability: Use an LLM and machine learning to process open-ended responses and identify inconsistencies.

Increase Understanding: Enable real-time questions regarding criteria and trait relevance.

Accessibility: Use plain language and high-contrast visuals to reduce cognitive load.

Scoring System:

The chatbot questions are mapped to the Diagnostic and Statistical Manual for Mental Disorders (DSM-5) criteria for autism, ADHD, OCD, and GAD.

Likert-scale responses are normalized to a value between 0 and 1. Open-ended responses are evaluated using OpenAI, which assigns a value between 0 and 1 based on how strongly the response endorses the diagnostic trait associated with the question.

Each normalized score is then weighted within its criterion by response type (presence, example, impact, and frequency) to prioritize clinically significant traits.

Criterion scores are then compared to predefined thresholds to determine whether each diagnostic criterion was met, borderline, or not met.

These criterion-level results are then evaluated according to the DSM-5 diagnostic structure to determine the final classification for each disorder.

Data Collection:

An anonymous survey asking about experiencing traits related to autism, ADHD, OCD, and GAD was completed by 14 participants between the ages of 13-17 after obtaining Informed Consent. See Figure 1 for Participant Composition.

Scoring Classification Testing:

Profile Testing: 10 "profiles" (simulated response sets consisting of two controls and two sets per disorder) were manually input into the tool. The tool’s results were compared to the intended classification.

Participant Testing: Utilized the responses collected through the Data Collection methodology. Each set of responses was manually input into the tool. The tool’s results were compared against the participants’ self-disclosed diagnoses.

Inconsistency Detection Testing:

Phase 1: 110 simulated response cases tested four types of inconsistencies: contradiction, contextual conflict, minimization, and insight-related. Each case was labelled obvious, subtle, or none to measure the system’s sensitivity. See Figure 2 for an example of a test case.

Phase 2: 12 simulated user profiles were created to evaluate inconsistency detection across a full questionnaire. Each profile contained 0–3 intentional inconsistencies.

Tool Architecture

The tool uses a structured, multi-stage screening process designed to evaluate whether experiences align with diagnostic traits associated with autism, ADHD, OCD, and GAD. The chatbot first asks questions regarding core diagnostic criteria. If the user endorses these experiences or challenges, the chatbot asks a follow-up set of questions specific to that disorder. The tool also prompts for clarification if a response is flagged as irrelevant or unclear.

A LLM (OpenAI API) analyzes open-ended responses for relevance and identifies potential inconsistencies. A basic rule-based system and a Random Forest machine learning model further evaluate whether detected inconsistencies are likely meaningful based on overall response patterns.

After all questions have been asked, the tool applies a weighted scoring system based on DSM-5 diagnostic criteria. Each score represents how strongly a user’s responses align with each criterion. The system then evaluates whether the overall scores meet the predefined thresholds for each disorder. This approach combines structured scoring with natural language analysis, enabling more nuanced evaluation than traditional static screening questionnaires.

The final results are presented in a clear and structured format to ensure user understanding. These results include:

A summary of the user’s responses

Any potential inconsistencies detected

Which disorders may align with their experiences

A recommendation regarding whether the user may wish to consider seeking a formal assessment.

See Figure 3 for the System Overview Diagram.

Inconsistency Testing Results

Overall, inconsistency detection testing demonstrated strong detection performance, with the system reliably identifying inconsistencies while minimizing false positives.

As shown in Table 2, the tool achieved an overall accuracy of 94.54% across 110 test cases in phase one. Direct contradictions and insight-related inconsistencies showed near-perfect detection (96.77% and 100% accuracy). Minimization and contextual cases achieved slightly lower accuracy (88.89% and 92%) due to three false negatives, but maintained high precision (94.12% and 94.44%). See Table 1 for phase one results by inconsistency type.

The tool achieved 100% precision and 87.5% accuracy across 12 simulated user profiles, as shown in Table 3. These results validate phase two’s objective: ensuring the inconsistency detection system does not flag excessive consistent responses.

Scoring Testing Results

Scoring classification testing indicates that the tool’s scoring system provides reliable classification of disorder alignment across both simulated profiles and real participant responses.

As shown in Table 5, the tool achieved 96.3% overall accuracy across simulated profiles. It demonstrated perfect classification for ADHD and OCD, with very high accuracy for autism (95%) and GAD (90%). Only two "adjacent matches" occurred (“may align” classification), confirming the scoring engine’s effectiveness in controlled scenarios. See Table 4 for a breakdown of participant scoring testing.

As shown in Table 6, the tool maintained a 92.86% overall accuracy when applied to real participant responses. The system achieved its highest reliability in identifying GAD (96.43%), followed by autism and OCD (92.86%). There were eight "adjacent matches" (“may align” classifications), 5 slight false positives and 3 slight false negatives.

Discussion:

Scoring and inconsistency testing highlight the tool’s accuracy in detecting inconsistencies within a user’s responses and identifying autism, ADHD, OCD, and GAD.

More false negatives than false positives reflects the design, which intentionally avoids over-flagging responses. This decision was made as missing some inconsistencies is less harmful than incorrectly flagging consistent responses. As a result, the system prioritizes precision over recall, meaning detected inconsistencies are highly reliable, though some may go undetected.

Most errors in phase one inconsistency detection testing were minimization cases and contextual inconsistencies, which involve nuanced language and tone. This requires the system to infer relationships between answers rather than identify explicit contradictions, a known limitation with AI. Despite this, the system successfully identified the majority of these inconsistencies.

In participant scoring classification testing, eight adjacent matches (“may align” classifications) occurred. Some may be attributed to variability in participant responses and differences between self-identification and endorsed traits. Thus, participant responses may sometimes partially match diagnostic criteria even when individuals do not identify with the disorder.

Conclusion:

In summary, barriers prevent many from accessing a formal diagnosis. For many individuals, self-diagnosis is the only option. However, self-diagnosed individuals often experience challenges such as imposter syndrome and self-doubt. The use of open-ended questions allows users to elaborate, prevents misunderstandings and does not force users into rigid answer choices. Overall, this tool shows strong potential as an accessible, interactive, and informative screening resource for individuals exploring whether their experiences may align with autism, ADHD, OCD, or GAD.

Future Testing:

Expand sample size to improve scoring validity.

Compare tool results with formal clinician assessments to evaluate alignment with professional diagnostic outcomes.

Future Improvements:

Obtain clinician feedback to refine and validate the chatbot’s questions.

Improve the inconsistency detection system by expanding ML training datasets, further integrating ML, and refining LLM prompts

Modify the tool to implement dynamic branching to immediately distinguish between potential disorders based on user input.

I would like to sincerely thank my dad for his support as my Adult Sponsor, my mom for her invaluable feedback throughout this project, and Katie Brent for her help editing and reviewing my project. Their guidance and encouragement were instrumental in completing this work.

Aguiar, Y. P., Galy, E., Godde, A., Trémaud, M., & Tardif, C. (2020). AutismGuide: A usability guidelines to design software solutions for users with autism spectrum disorder.

Behaviour & Information Technology, 41(6), 1132–1150. https://doi.org/10.1080/0144929x.2020.1856927

Engelbrecht, N. (2021, July 21). Decoding autism in the DSM-5. Embrace Autism. https://embrace-autism.com/decoding-autism-in-the-dsm-5/

Centre for Addiction and Mental Health. (n.d.). Adult Neurodevelopmental Services (ANS). https://www.camh.ca/-/media/files/ans-faqs-pdf.pdf

Goldenbaum, B. (2025, January 23). AI-powered Chatbots: Accessibility best practices you need to know. A2J Tech.

Hours, C., Recasens, C., & Baleyte, J.-M. (2022). ASD and ADHD comorbidity: What are we talking about? Frontiers in Psychiatry, 13. https://doi.org/10.3389/fpsyt.2022.837424

Kentrou, V., Livingston, L. A., Grove, R., Hoekstra, R. A., & Begeer, S. (2024). Perceived misdiagnosis of psychiatric conditions in autistic adults. eClinicalMedicine, 71, 102586.  https://doi.org/10.1016/j.eclinm.2024.102586

Lewis, L. F. (2019). Barriers to formal diagnosis of autism spectrum disorder in adults.

Encyclopedia of Autism Spectrum Disorders, 1–6. https://doi.org/10.1007/978-1-4614-6435-8_102410-1

Marsh, C. (2025). Autism Identification Tool: An AI Approach to Reducing Diagnostic Barriers. ProjectBoard. https://partner.projectboard.world/ysc/project/autism-identification-tool-an-ai-approach-to-reducing-diagnostic-barriers

Ratto, A. B., Bascom, J., daVanport, S., Strang, J. F., Anthony, L. G., Verbalis, A., Pugliese, C., Nadwodny, N., Brown, L. X. Z., Cruz, M., Hector, B. L., Kapp, S. K., Giwa Onaiwu, M., Raymaker, D. M., Robison, J. E., Stewart, C., Stone, R., Whetsell, E., Pelphrey, K., & Kenworthy, L. (2023). Centering the inner experience of autism: Development of the self-assessment of autistic traits. Autism in Adulthood, 5(1), 93–105. https://doi.org/10.1089/aut.2021.0099

Sohn, J.-S., Lee, E., Kim, J.-J., Oh, H.-K., & Kim, E. (2025). Implementation of generative AI for the assessment and treatment of autism spectrum disorders: A scoping review.

Frontiers in Psychiatry, 16. https://doi.org/10.3389/fpsyt.2025.1628216

The Other Autism. (2023, August 16). What Is Autistic Impostor Syndrome? https://other-autism.com/2023/08/16/what-is-autistic-impostor-syndrome/

Words Potrayal. (n.d.). [Picture of cartoon laptop] [Illustration]. Vecteezy. https://www.vecteezy.com/vector-art/27887778-kids-drawing-cartoon-vector-illustration-laptop-icon-isolated-on-white-background

Zaboski, B. A., & Storch, E. A. (2018). Comorbid Autism Spectrum Disorder and anxiety

disorders: A brief review. Future Neurology, 13(1), 31–37. https://doi.org/10.2217/fnl-2017-0030