Pediatric Autism Diagnosis Accuracy and Confidence: A Comparison of Experienced and Inexperienced Clinicians Making Decisions with and without AI Decision Support

Technical Summary: Pediatric Autism Diagnosis Accuracy and Confidence with AI Decision Support

Problem Statement
Early diagnosis of Autism Spectrum Disorder (ASD) is critical for optimal intervention, yet diagnosis is often delayed due to a shortage of clinicians with ASD expertise. While Artificial Intelligence (AI) has shown promise in identifying behavioral patterns associated with ASD, most research focuses on algorithmic accuracy under controlled conditions rather than clinical integration. A significant gap exists in understanding how AI decision support tools affect the diagnostic accuracy and confidence of clinicians with varying levels of experience, particularly regarding the risks of automation bias (over-reliance on automated outputs) and the "black box" nature of many AI systems.

Methodology
The study evaluated Autism Diagnostic Identification Software (ADIS), a clinical decision support tool designed to provide pre-decision support. Unlike Explainable AI (XAI) frameworks that offer post-hoc explanations, ADIS highlights autism-relevant behaviors in narrative clinical text (from Electronic Health Records or the CDC's ADDM Network) and labels them according to DSM-5 criteria using a BioBERT model. The system suggests a diagnosis based on the count of labeled behaviors and allows clinicians to adjust the strictness of the labeling via a probability threshold slider.

The researchers conducted a vignette-based user study with 21 clinicians (48% with completed medical training, 52% trainees). Participants reviewed four real pediatric cases each: two with ADIS active and two without. The study design intentionally included cases where ADIS generated correct and incorrect diagnostic suggestions to test how clinicians responded to AI errors. Participants were asked to decide whether to refer a child for autism evaluation and to rate their confidence. The study also collected qualitative data on tool usability and the nature of the evidence participants provided for their decisions.

Key Contributions

• Pre-Decision Support Model: The study introduces and evaluates a workflow where AI highlights specific behavioral evidence before a clinician makes a diagnostic decision, rather than offering a final diagnosis or post-hoc explanation.
• Interaction of Expertise and AI Accuracy: The research explicitly isolates the impact of AI correctness on clinicians with different levels of training, revealing distinct reliance patterns between trainees and experienced clinicians.
• Usability of Transparent AI: It demonstrates that a transparent, rule-based AI system (linking text snippets to DSM-5 criteria) can be perceived as usable and helpful, even when it is not perfect.

Results

• Diagnostic Accuracy: Overall accuracy was 57.14% when ADIS was active versus 66.67% without ADIS, a non-significant difference. However, a significant interaction emerged based on training and AI correctness:
  • Trainees (No completed medical degree): Achieved 0% accuracy when ADIS was wrong but 68.75% accuracy when ADIS was correct.
  • Experienced Clinicians (Completed medical degree): Achieved 66.67% accuracy when ADIS was wrong (showing skepticism/correction) but only 56.25% accuracy when ADIS was correct (suggesting they sometimes ignored valid AI signals).
• Decision Confidence: Confidence was significantly higher when ADIS was active (mean score 1.33 vs. 1.09, p = .058 trend) and significantly higher among trained clinicians. Crucially, confidence remained high regardless of whether the AI suggestion was correct or incorrect, indicating a decoupling of confidence from accuracy.
• Usability: 90% of participants felt the tool could be learned quickly. Most rated the features as well-integrated and found the tool easy to use, though some requested clearer guidance on the "strict vs. lenient" threshold settings.
• Communication Patterns: Trained clinicians provided shorter, more concise responses to open-ended questions and used DSM-focused language (27.7%) significantly more often than trainees (8.7%), who provided more detailed but less formally structured descriptions.

Significance and Claims
The paper claims that while AI support can augment the decision-making of inexperienced clinicians, it introduces a significant risk of automation bias, particularly when the AI is incorrect. The study highlights that:

1. AI Accuracy is Paramount: The utility of the tool is heavily dependent on the correctness of the AI; when the AI was wrong, trainees failed to detect the error, leading to a 0% success rate in those specific instances.
2. Confidence Inflation: The presence of AI increases clinician confidence even when the AI is wrong, posing a patient safety concern.
3. Training Necessity: The findings suggest that deploying such tools requires AI literacy training to mitigate automation bias. The authors argue that interpretability (showing the evidence) is necessary but not sufficient; clinicians must be trained to calibrate their trust in the system.
4. Expertise Gap: There is a distinct difference in how trainees and experts interact with AI. Trainees tend to uncritically accept AI outputs, while experts may exhibit "productive skepticism" that sometimes leads to rejecting valid AI signals.

The authors conclude that ADIS is a usable and helpful tool, but its effective deployment in clinical practice depends on pairing it with education that helps clinicians, especially trainees, distinguish between valid AI signals and errors, thereby preventing the over-reliance that undermines diagnostic accuracy.