Expertise-Dependent Verification Behaviors in Response to AI Reliability in Neonatal Pneumoperitoneum

Lee et al. · Manuscript in Preparation

One-line takeaway: AI reliability does not translate linearly into clinical benefit. In a multi-reader, multi-center study, we demonstrate that AI reliability acts as a modulator of verification behavior. While "Reliable" AI induced automation bias in trainees, "Unreliable" AI paradoxically triggered sentinel (vigilant) behavior in experts.

---

🖼️ Visual Abstract: The Sentinel Effect vs. Automation Bias

Visualizing the study's core behavioral discovery:

This illustration represents the divergent verification behaviors observed in our multi-center study:

• Left (Automation Bias): When interacting with Reliable AI (Blue stream), less experienced readers (Residents) tended to "offload" cognition. The smooth, accurate advice paradoxically reduced vigilance, leading to the acceptance of errors when they did occur.
• Right (The Sentinel Effect): When interacting with Error-Injected AI (Red/Glitchy stream), experts (Neonatologists) engaged a "cognitive forcing function." The presence of systematic AI errors triggered a high-vigilance mode, where experts actively used tools to refute the AI, acting as a safety shield for the patient.

Generated via NanoBanana based on the research paper's findings.

👾 Model Architecture

To ensure that differences in reader behavior were driven solely by reliability (and not model capacity), both the "Reliable" and "Error-Injected" assistants utilized the exact same underlying architecture:

• Backbone: RAD-DINO (ViT-B/14), a vision foundation model pre-trained on large-scale radiology datasets.
• Adaptation (LoRA): Parameter-efficient fine-tuning using Low-Rank Adaptation ($r=12, \alpha=24$) injected into Query/Value projections and the MLP layer. Only 1.36% of parameters were trainable, preventing catastrophic forgetting.
• Sampling Strategy (RFBS): A custom Representation-Focused Batch Sampler was used to enforce diversity and ensure exposure to uncommon pneumoperitoneum distributions during training.

Note: The "Error-Injected" model was created using this same robust architecture but trained on systematically poisoned labels (e.g., labeling drains or pneumatosis as free air) to simulate "plausible" AI errors.

---

🎯 Why This Matters

Neonatal pneumoperitoneum is a time-critical surgical emergency. However, integrating AI into this high-stakes workflow is not just about model accuracy (AUC). Clinicians interact with advice, confidence cues, and time pressure.

This study investigates the Human-AI Interaction (HAI) layer:

1. Verification Phenotypes: When AI is highly capable, does it help—or does it reduce human effort?
2. The "Sentinel" Effect: When AI is systematically wrong, do clinicians disengage, blindly follow, or become hyper-vigilant?
3. Expertise Gradient: Do neonatologists, radiologists, and residents react differently to the same AI signals?

---

🔬 Study Overview

1. Cohorts

• Internal Development: 688 radiographs (310 positive) from 216 neonates.
• External Validation (Reader Study): 125 radiographs (40 positive) from 11 tertiary hospitals via AI-Hub.

2. The Reader Study

• Participants (N=14):
  • Pediatric Radiologists ($n=6$)
  • Neonatologists ($n=3$)
  • Radiology Residents ($n=5$)
• Design: Two-session, counterbalanced MRMC Crossover with a 6-week washout.

• Case Allocation (Stratified):
  • Unaided ($n=41$)
  • Reliable AI ($n=40$)
  • Error-Injected AI ($n=44$)

Note: Reliability was fixed at the case level. Readers were blind to the specific reliability condition of any given case, ensuring natural behavioral responses.

---

🤖 AI Tools Evaluated

We compared two distinct "advice streams" generated by the same Vision Foundation Model (RAD-DINO + LoRA) architecture:

Model Phenotype	AUC (Study Subset)	Characteristics	Purpose
Reliable AI	0.861	High-performance, standard training.	To test Automation Complacency.
Error-Injected AI	0.43	Intentionally engineered via systematic label bias (e.g., drains, pneumatosis).	To test Sentinel Behavior (response to failure).

⚠️ Clinical Note: The Error-Injected AI is an experimental probe only and is not intended for clinical deployment.

---

📊 Statistical Methodology

Our primary analysis utilizes a Crossed Random-Effects Generalized Linear Mixed Model (GLMM).

Unlike simple aggregate comparisons, this method:

1. Controls for Case Difficulty: Random intercept for Case_ID ($1 | filename$) accounts for intrinsic image difficulty.
2. Controls for Reader Variability: Random intercept for Reader_ID ($1 | reader$) accounts for individual competence.
3. Isolates Interaction: Accurately estimates the Condition × Expertise interaction term.

---

🗝️ Key Findings

1. The Interaction Effect: Divergent Trajectories

The primary GLMM revealed a significant interaction ($p=0.020$) between Neonatologists and the Error-Injected AI condition.

Contrast (vs Pediatric Radiologist)	Odds Ratio (OR)	95% CI	P-value
Interaction: Error-Injected × Neonatologist	4.16	1.26–13.77	0.020

Interpretation: While the reference group (Radiologists) showed stable or declining odds under error-prone AI, Neonatologists exhibited a statistically distinct positive trajectory, indicating a shift toward active verification ("Sentinel Behavior").

2. Mechanisms: Trust vs. Vigilance

We decomposed behavior in cases where the AI was incorrect:

• Automation Bias (Residents): Rapidly accepted incorrect advice from "Reliable" AI in 52.0% of discordant cases.
• Sentinel Behavior (Neonatologists): Successfully overrode "Error-Injected" AI in 91.6% of discordant cases.

3. Verification Effort (Time Analysis)

Disagreement acted as a Cognitive Forcing Function for experts:

• Neonatologists: 10.04s (Discordant) vs 5.42s (Concordant) → +4.62s deliberation penalty.
• Residents: No significant time penalty ($p>0.05$), consistent with "cognitive short-circuiting" (rapid acceptance/rejection without deep verification).

4. Explainability (Saliency Maps)

• Residents: Viewed CAMs frequently (53.8% of errors) but accuracy remained low (78.2%).
• Experts: Viewed CAMs sparingly (17.2%) but achieved 100% accuracy when they did, suggesting "utility-triggered" verification.

---

📜 Citation

If you use these findings or the "Error-Injection" validation framework, please cite:

Lee et al. Expertise-Dependent Verification Behaviors in Response to AI Reliability in Neonatal Pneumoperitoneum: A Multi-Reader Multi-Center Study. (Under Review, 2026)

---

✉️ Correspondence

• Namkug Kim, PhD (Medical Imaging and Intelligent Reality Lab, Asan Medical Center)
• Hee Mang Yoon, MD, PhD (Massachusetts General Hospital / Asan Medical Center)