During the COVID-19 pandemic, staying home proved to be an effective way to mitigate the spread of the virus. Stay-at-home orders and guidelines were issued by governments across the globe and were followed by a large portion of the population in the early stages of the outbreak when there was a lack of COVID-specific medical knowledge. The decision of whether to stay home came with many trade-offs, such as risking personal exposure to the virus when leaving home or facing financial and mental health burdens when remaining home. In this research, we study how individuals make strategic decisions to balance these conflicting outcomes. We present a model to study individuals’ decision making based on decision and prospect theory, and we conduct sensitivity analysis to study the fluctuations in optimal strategies when there are changes made to the model’s parameters. A Monte Carlo simulation is implemented to further study the performance of our model, and we compare our simulation results with real data that captures individuals’ stay-at-home decisions. Overall, this research models and analyzes the behaviors of individuals during the COVID-19 pandemic and can help support decision making regarding control measures and policy development when public health emergencies appear in the future. History: This article was accepted for the Decision Analysis Special Issue on Emerging Topics in Health. Funding: The first two authors’ efforts were supported by the National Natural Science Foundation of China [Grants 71971217 and 71671189], the Key Project of Natural Science Foundation of China [Grants 71790615 and 91846301], and the Independent Exploration of Innovation Project for Postgraduate of Central South University [Grant 2019zzts843]. The third author’s effort was partially supported by the U.S. National Science Foundation Graduate Research Fellowship Program [Grant 2043091].

Background: Large language models (LLMs) are increasingly used by patients for health information and preliminary medical advice. In patient-facing consultations, users may present explicitly stated diagnostic preferences or symptom narratives emphasizing a preferred explanation. Such cognitively biased input constrains the diagnostic context available to the model and may systematically steer its reasoning during interactive LLM-supported health consultations. Objective: This study aimed to quantify the impact of patient cognitive bias on LLM diagnostic performance in multiturn consultations, assess the effectiveness of prompt-based mitigation strategies and decoding temperature adjustment, and evaluate a dual-system framework for improving robustness under biased interaction. Methods: We developed a simulated patient agent to generate both unbiased and cognitively biased consultations using 1273 medical question answering dataset United States Medical Licensing Examination cases. Six widely used LLMs of varying capacities were evaluated through 3-round, multiturn dialogues, after which each model produced a final diagnostic judgment based on the complete consultation record. Diagnostic accuracy was the primary outcome. Secondary outcomes included bias-induced accuracy decline (absolute reduction in accuracy under biased vs standard consultations) and bias-influenced error proportion (proportion of incorrect responses aligned with the patient’s preferred but incorrect diagnosis). Three prompt-based mitigation strategies and 4 decoding temperature settings were tested. In addition, a dual-system framework was evaluated, in which a conversational foundation LLM conducted patient interaction and history taking (System 1), while a reasoning-oriented LLM (o1-mini) generated the final diagnostic judgment (System 2). In the foundation-only condition, the same LLM performed both interaction and diagnosis. Results: Across all 6 evaluated models, cognitively biased consultations led to marked diagnostic accuracy declines of approximately 7 to 39 percentage points compared with standard multiturn consultations, whereas static single-response tests and standard consultations showed comparable accuracy. Larger deteriorations were observed in lower-capacity models, with some approaching random-guess performance under bias. Errors were frequently aligned with patient bias, with bias-influenced error proportion exceeding one-third across models, indicating systematic conformity rather than random error. Prompt-based mitigation strategies and decoding temperature reduction yielded limited and inconsistent improvements and did not reliably prevent bias-induced performance loss. By contrast, the dual-system framework substantially improved diagnostic accuracy under biased conditions, producing gains of approximately 10 to 39 percentage points across most models and recovering a large proportion of the performance lost due to bias, particularly in lower-capacity systems. Conclusions: Patient-driven cognitive bias represents an underrecognized behavioral risk in LLM-supported health consultations. Common mitigation approaches, such as prompt engineering or decoding parameter adjustment, provide limited resilience. Explicitly separating conversational interaction from deliberative diagnostic reasoning through a dual-system framework enables more robust diagnostic performance under biased input while potentially preserving patient-facing dialogue fluency by retaining the foundation LLM as the conversational component, offering a scalable design strategy for safer medical AI systems.

Effective communication is essential in clinical training, yet opportunities for realistic and interactionally authentic practice remain limited. Standardized patients (SPs) provide realism but are resource-intensive, whereas virtual patients (VPs) offer scalability but have limited capacity to reproduce responsive, interactionally authentic dialogue. Recent advances in generative artificial intelligence (AI), particularly real-time voice models, have opened new possibilities for natural and synchronous dialogue in virtual simulations. To evaluate whether a real-time voice-based virtual patient (RT-VP) can achieve communication performance, perceived realism, and self-efficacy outcomes comparable to SP training, thereby addressing the global challenge of scalable, authentic communication training. The RT-VP was developed on the Doubao real-time voice generative AI platform, which supports synchronous, bidirectional spoken interaction. In a randomized controlled study, 134 residents were assigned to RT-VP, standardized-patient (SP), or peer role-play (PR) groups. All groups received identical SPIKES-based instruction and practiced in their assigned simulation modality. Outcomes included communication performance, self-efficacy, and perceived realism. Post-training SPIKES scores were highest for SP (26.6 ± 3.0), followed by RT-VP (24.9 ± 2.9) and PR (20.4 ± 4.3) (p < .001). Both SP and RT-VP outperformed PR (p < .001), and the difference between SP and RT-VP did not reach statistical significance (p = .06). Perceived realism generally followed an SP > RT-VP > PR pattern, with RT-VP demonstrating comparable linguistic realism to SP, while SP remained superior in contextual, emotional, and engagement realism. This study represents, to our knowledge, the first controlled comparison of real-time voice generative AI-based VPs and SP encounters in communication training. RT-VP simulation approximated SP-level communication performance and perceived authenticity while offering scalability and consistency. Real-time voice simulation offers a scalable means to expand access to emotionally authentic communication training worldwide.