Foundation Models that are capable of processing and generating multi-modal data have transformed AI’s role in medicine. However, a key limitation of their reliability is hallucination, where inaccurate or fabricated information can impact clinical decisions and patient safety. We define medical hallucination as any instance in which a model generates misleading medical content. This paper examines the unique characteristics, causes, and implications of medical hallucinations, with a particular focus on how these errors manifest themselves in real-world clinical scenarios. Our contributions include (1) a taxonomy for understanding and addressing medical hallucinations, (2) benchmarking models using medical hallucination dataset and physician-annotated LLM responses to real medical cases, providing direct insight into the clinical impact of hallucinations, and (3) a multi-national clinician survey on their experiences with medical hallucinations. Our results reveal that inference techniques such as Chain-of-Thought (CoT) and Search Augmented Generation can effectively reduce hallucination rates. However, despite these improvements, non-trivial levels of hallucination persist. These findings underscore the ethical and practical imperative for robust detection and mitigation strategies, establishing a foundation for regulatory policies that prioritize patient safety and maintain clinical integrity as AI becomes more integrated into healthcare. The feedback from clinicians highlights the urgent need for not only technical advances but also for clearer ethical and regulatory guidelines to ensure patient safety. A repository organizing the paper resources, summaries, and additional information is available at https://github.com/mitmedialab/medical_hallucination.

There is a growing body of research revealing that longitudinal passive sensing data from smartphones and wearable devices can capture daily behavior signals for human behavior modeling, such as depression detection. Most prior studies build and evaluate machine learning models using data collected from a single population. However, to ensure that a behavior model can work for a larger group of users, its generalizability needs to be verified on multiple datasets from different populations. We present the first work evaluating cross-dataset generalizability of longitudinal behavior models, using depression detection as an application. We collect multiple longitudinal passive mobile sensing datasets with over 500 users from two institutes over a two-year span, leading to four institute-year datasets. Using the datasets, we closely re-implement and evaluated nine prior depression detection algorithms. Our experiment reveals the lack of model generalizability of these methods. We also implement eight recently popular domain generalization algorithms from the machine learning community. Our results indicate that these methods also do not generalize well on our datasets, with barely any advantage over the naive baseline of guessing the majority. We then present two new algorithms with better generalizability. Our new algorithm, Reorder, significantly and consistently outperforms existing methods on most cross-dataset generalization setups. However, the overall advantage is incremental and still has great room for improvement. Our analysis reveals that the individual differences (both within and between populations) may play the most important role in the cross-dataset generalization challenge. Finally, we provide an open-source benchmark platform GLOBEM- short for Generalization of Longitudinal BEhavior Modeling - to consolidate all 19 algorithms. GLOBEM can support researchers in using, developing, and evaluating different longitudinal behavior modeling methods. We call for researchers' attention to model generalizability evaluation for future longitudinal human behavior modeling studies.

Although large language models (LLMs) show promise for clinical healthcare applications, their utility for personalized health monitoring using wearable device data remains underexplored. Here we introduce the Personal Health Large Language Model (PH-LLM), designed for applications in sleep and fitness. PH-LLM is a version of the Gemini LLM that was finetuned for text understanding and reasoning when applied to aggregated daily-resolution numerical sensor data. We created three benchmark datasets to assess multiple complementary aspects of sleep and fitness: expert domain knowledge, generation of personalized insights and recommendations and prediction of self-reported sleep quality from longitudinal data. PH-LLM achieved scores that exceeded a sample of human experts on multiple-choice examinations in sleep medicine (79% versus 76%) and fitness (88% versus 71%). In a comprehensive evaluation involving 857 real-world case studies, PH-LLM performed similarly to human experts for fitness-related tasks and improved over the base Gemini model in providing personalized sleep insights. Finally, PH-LLM effectively predicted self-reported sleep quality using a multimodal encoding of wearable sensor data, further demonstrating its ability to effectively contextualize wearable modalities. This work highlights the potential of LLMs to revolutionize personal health monitoring via tailored insights and predictions from wearable data and provides datasets, rubrics and benchmark performance to further accelerate personal health-related LLM research.

Passively collected behavioral health data from ubiquitous sensors could provide mental health professionals valuable insights into patient's daily lives, but such efforts are impeded by disparate metrics, lack of interoperability, and unclear correlations between the measured signals and an individual's mental health. To address these challenges, we pioneer the exploration of large language models (LLMs) to synthesize clinically relevant insights from multi-sensor data. We develop chain-of-thought prompting methods to generate LLM reasoning on how data pertaining to activity, sleep and social interaction relate to conditions such as depression and anxiety. We then prompt the LLM to perform binary classification, achieving accuracies of 61.1%, exceeding the state of the art. We find models like GPT-4 correctly reference numerical data 75% of the time. While we began our investigation by developing methods to use LLMs to output binary classifications for conditions like depression, we find instead that their greatest potential value to clinicians lies not in diagnostic classification, but rather in rigorous analysis of diverse self-tracking data to generate natural language summaries that synthesize multiple data streams and identify potential concerns. Clinicians envisioned using these insights in a variety of ways, principally for fostering collaborative investigation with patients to strengthen the therapeutic alliance and guide treatment. We describe this collaborative engagement, additional envisioned uses, and associated concerns that must be addressed before adoption in real-world contexts.

Deriving personalized insights from popular wearable trackers requires complex numerical reasoning that challenges standard LLMs, necessitating tool-based approaches like code generation. Large language model (LLM) agents present a promising yet largely untapped solution for this analysis at scale. We introduce the Personal Health Insights Agent (PHIA), a system leveraging multistep reasoning with code generation and information retrieval to analyze and interpret behavioral health data. To test its capabilities, we create and share two benchmark datasets with over 4000 health insights questions. A 650-hour human expert evaluation shows that PHIA significantly outperforms a strong code generation baseline, achieving 84% accuracy on objective, numerical questions and, for open-ended ones, earning 83% favorable ratings while being twice as likely to achieve the highest quality rating. This work can advance behavioral health by empowering individuals to understand their data, enabling a new era of accessible, personalized, and data-driven wellness for the wider population.

Large language models (LLMs) have shown remarkable abilities to generate code. However, their ability to develop software for physical computing and embedded systems, which requires cross-domain hardware and software knowledge, has not been thoroughly studied. We observe through our experiments and a 15-user pilot study that even when LLMs fail to produce working code, they can generate helpful reasoning about embedded design tasks, as well as specific debugging suggestions for both novice and expert developers. These results highlight the potential to develop AI assistants to dramatically lower the barrier to entry for working with hardware. To evaluate the capabilities and limitations of LLMs, we develop an automated testbench to quantify LLM performance on embedded programming tasks and perform 450 trials. We leverage these findings to analyze how programmers interact with these tools including their productivity and sense of fulfillment and outline a human-AI collaborative workflow for developing and debugging embedded systems.