JMIR Publications

ABSTRACT

Background:

Emergency department (ED) triage relies on visible indicators and can miss occult injuries such as delayed epidural hemorrhage. Patient digital twins that integrate brain–computer interface (BCI) signals and deep biosignals can detect such occult deterioration in real time and recommend a severity upgrade; however, these recommendations frequently collide with clinicians’ heuristics and elicit algorithm aversion. How a clinician makes sense of, and ultimately accepts or rejects, a counter-heuristic artificial intelligence (AI) alert is a central cognitive-ergonomics and decision-support problem that remains poorly understood.

Objective:

This study examined how the explanation modality of an explainable AI (XAI) decision aid (causal narrative vs counterfactual) and clinician expertise (novice vs expert) jointly shape epistemic trust, the sensemaking trajectory, and triage reassessment in an ED digital twin, and it derives an expertise-adaptive human–machine interface (HMI) design.

Methods:

We conducted an in-silico persona experiment in which a large language model (LLM) was prompted with emergency-medicine novice and expert personas. Using a 2 (expertise) × 2 (explanation modality) full-factorial design with 30 replications per cell (N=120 Chain-of-Thought reasoning traces), we combined quantitative indicators (acceptance rate, number of reasoning steps to conflict resolution, and frequency of epistemic-trust markers) with constructivist grounded-theory coding. Personas were specified by expertise and experience only—cognitive bias was treated as an observed outcome rather than an injected instruction—the stimulus was held constant, and the base model, temperature (0.7), and per-run seed rotation were fixed so that the 120 traces represent a response distribution to an identical stimulus.

Results:

A robust expertise × explanation interaction emerged. For experts, counterfactual explanations raised acceptance markedly relative to causal narratives (50.0% vs 76.7%) and shortened conflict resolution (6.5 vs 4.4 reasoning steps), whereas for novices the two modalities differed little (86.7% vs 80.0%). Three core categories organized the sensemaking process: clinical heuristic dominance, sensemaking reconfiguration, and cognitive conflict-resolution strategies. Counterfactual explanations appeared to disrupt experts’ confirmation loop (“what would change if I ignored this signal?”), whereas novices showed high but potentially uncritical acceptance.

Conclusions:

In high-stakes ED decision support, algorithm aversion is better understood as a problem of mind perception and the redistribution of agency than as a fixed trust trait. XAI explanation modality operates as a cognitive-ergonomic lever whose effect is moderated by expertise, motivating expertise-adaptive HMI and autonomy-preserving governance. Because the evidence is generated in silico, the quantitative estimates should be read as a structured set of hypotheses for subsequent human validation rather than as confirmatory human findings. Clinical Trial: Not applicable. This study is an in-silico simulation using large language models and does not involve human participants or actual patient interventions.