Benjamin W. Nelson

Large language models often mishandle psychiatric emergencies, offering harmful or inappropriate advice. This study evaluated the Verily Mental Health Guardrail (VMHG) on two clinician-labeled datasets: the Verily Mental Health Crisis Dataset v1.0, containing 1800 simulated messages and the NVIDIA Aegis AI Content Safety Dataset subsetted to 794 mental health-related messages. Performance was benchmarked against OpenAI Omni Moderation Latest and NVIDIA NeMo Guardrails. The VMHG demonstrated high sensitivity (0.990) and specificity (0.992) on the Verily dataset, with an F1-score of 0.939 and high category-level sensitivity (0.917–0.992) and specificity (≥0.978). On the NVIDIA dataset, it maintained strong sensitivity (0.982) and accuracy (0.921) with reduced specificity (0.859). Compared with NVIDIA and OpenAI guardrails, the VMHG achieved significantly higher sensitivity (all p < 0.001) and comparable specificity (NVIDIA p < 0.001, OpenAI p = 0.094). Overall, the VMHG demonstrated robust, generalizable, and clinically oriented safety performance that prioritizes sensitivity to minimize missed mental health crises.

Journal of Medical Internet Research - International Scientific Journal for Medical Research, Information and Communication on the Internet #Preprint #PeerReviewMe: Warning: This is a unreviewed preprint. Readers are warned that the document has not been peer-reviewed by expert/patient reviewers or an academic editor, may contain misleading claims, and is likely to undergo changes before final publication, if accepted, or may have been rejected/withdrawn. Readers with interest and expertise are encouraged to sign up as peer-reviewer, if the paper is within an open peer-review period. Please cite this preprint only for review purposes or for grant applications and CVs (if you are the author). Background: Health care chatbots are rapidly proliferating, while generative artificial intelligence (AI) outpaces existing evaluation standards. Objective: We aimed to develop a structured, stakeholder-informed framework to standardize evaluation of health care chatbots. Methods: PRISMA (Preferred Reporting Items for Systematic reviews and Meta-Analyses)–guided searches across multiple databases identified 266 records; 152 were screened, 21 full texts were assessed, and 11 frameworks were included. We extracted 356 questions (refined to 271 by deduplication and relevance review), mapped items to Coalition for Health AI constructs, and organized them with iterative input from clinicians, patients, developers, epidemiologists, and policymakers. Results: We developed the Health Care AI Chatbot Evaluation Framework (HAICEF), a hierarchical framework with 3 priority domains (safety, privacy, and fairness; trustworthiness and usefulness; and design and operational effectiveness) and 18 second-level and 60 third-level constructs covering 271 questions. Emphasis includes data provenance and harm control; Health Insurance Portability and Accountability Act/General Data Protection Regulation–aligned privacy and security; bias management; and reliability, transparency, and workflow integration. Question distribution across domains is as follows: design and operational effectiveness, 40%; trustworthiness and usefulness, 39%; and safety, privacy and fairness, 21%. The framework accommodates both patient-facing and back-office use cases. Conclusions: HAICEF provides an adaptable scaffold for standardized evaluation and responsible implementation of health care chatbots. Planned next steps include prospective validation across settings and a Delphi consensus to extend accountability and accessibility assurances.

We evaluated the ability of three leading LLMs (GPT-4o, Gemini 2.0 Experimental, and Claude 3.5 Sonnet) to recognize human facial expression using the NimStim dataset. GPT and Gemini matched or exceeded human performance, especially for calm/neutral and surprise. All models showed strong agreement with ground truth, though fear was often misclassified. Findings underscore the growing socioemotional competence of LLMs and their potential for healthcare applications.

Large language models often mishandle psychiatric emergencies, offering harmful or inappropriate advice and enabling destructive behaviors. This study evaluated the Verily behavioral health safety filter (VBHSF) on two datasets: the Verily Mental Health Crisis Dataset containing 1,800 simulated messages and the NVIDIA Aegis AI Content Safety Dataset subsetted to 794 mental health-related messages. The two datasets were clinician-labelled and we evaluated performance using the clinician labels. Additionally, we carried out comparative performance analyses against two open source, content moderation guardrails: OpenAI Omni Moderation Latest and NVIDIA NeMo Guardrails. The VBHSF demonstrated, well-balanced performance on the Verily Mental Health Crisis Dataset v1.0, achieving high sensitivity (0.990) and specificity (0.992) in detecting any mental health crises. It achieved an F1-score of 0.939, sensitivity ranged from 0.917-0.992, and specificity was >= 0.978 in identifying specific crisis categories. When evaluated against the NVIDIA Aegis AI Content Safety Dataset 2.0, VBHSF performance remained highly sensitive (0.982) and accuracy (0.921) with reduced specificity (0.859). When compared with the NVIDIA NeMo and OpenAI Omni Moderation Latest guardrails, the VBHSF demonstrated superior performance metrics across both datasets, achieving significantly higher sensitivity in all cases (all p < 0.001) and higher specificity relative to NVIDIA NeMo (p < 0.001), but not to OpenAI Omni Moderation Latest (p = 0.094). NVIDIA NeMo and OpenAI Omni Moderation Latest exhibited inconsistent performance across specific crisis types, with sensitivity for some categories falling below 0.10. Overall, the VBHSF demonstrated robust, generalizable performance that prioritizes sensitivity to minimize missed crises, a crucial feature for healthcare applications.

Smartphones are essential tools in daily life yet concerns persist about their potential effects on mental health and well-being. Research on these relationships often relies on subjective measures of smartphone use, which can lack validity. These studies are also often limited by small, homogenous samples and the inability to differentiate between types of smartphone activities (e.g., social media vs. other applications). In this study, we analyzed 250,000 d of smartphone usage data from 10,099 adults, using objective measures of smartphone usage collected over 4 wk. We examined normative usage patterns and the temporal relationships between smartphone use and mood. Our findings reveal weak or null associations between smartphone use and mood, both cross-sectionally and longitudinally, and at both individual and group levels. Demographic factors, such as age and gender, were stronger predictors of mood than smartphone use. Statistically significant results that were observed showed that younger adults demonstrated a stronger association between social media use and lower mood cross-sectionally, but not longitudinally. Non-social app use was linked to lower mood in between-subjects analyses, but better mood in within-subjects analyses. However, all effects were small (or null), indicating minimal practical significance. These findings suggest that smartphone use has a limited or negligible impact on mood and well-being over a 4-wk period and do not support claims of a strong or causal relationship between smartphone use and mental well-being across this timeframe.

The goal was to evaluate the performance of a suite of 12 sleep measures generated by a multi‐sensor wrist‐worn wearable device, the verily numetric watch (VNW), in a diverse cohort. We used polysomnography (PSG) as reference during one‐night simultaneous recording in a sample of N = 41 (18 male, age range: 18–78 years). We performed epoch‐by‐epoch comparisons for all measures. Key specific analyses were: core accuracy metrics for sleep versus wake classification; bias for continuous measures (Bland–Altman); unweighted Cohen's kappa and accuracy for sleep stage classifications and mean count difference and linearly weighted Cohen's kappa for count metric. In addition, we performed exploratory subgroup analyses by sex, age, skin tone, body mass index and arm hair density. Sensitivity and specificity (95% CI) of sleep versus wake classification were 0.97 (0.96, 0.98) and 0.66 (0.61, 0.71), respectively. Mean total sleep time bias was 14.55 min (1.61, 27.16); wake after sleep onset, −11.77 min (−23.89, 1.09); sleep efficiency, 3.15% (0.68, 5.57); sleep onset latency, −3.24 min (−9.38, 3.57); light‐sleep duration, 3.78 min (−7.04, 15.06); deep‐sleep duration, 3.91 min (−4.59, 12.60) and rapid eye movement‐sleep duration, 6.94 min (0.57, 13.04). Mean difference for the number of awakenings, 0.17: (95% CI: −0.32, 0.71) and overall accuracy of sleep stage classification, 0.78 (0.51, 0.88). Most measures showed statistically significant proportional biases and/or heteroscedasticity. Exploratory subgroup results appeared largely consistent with the overall group, although small samples precluded strong conclusions. These results support the use of VNWs in classifying sleep versus wake, sleep stages and overnight sleep measures.

Background. Evaluating the social and emotional capabilities of large language models (LLMs), such as their ability to recognize human facial emotion is critical as their role in human-computer interactions (HCIs) expands, particularly in healthcare applications. Facial expressions convey affective and clinical information, useful for detecting emotions, contextualizing language, understanding interpersonal dynamics, and identifying potential mental health and neurocognitive disorders. However, LLMs' ability to accurately interpret facial expressions remains unclear.Methods. We evaluated the agreement and accuracy of three leading LLM models, GPT-4o, Gemini 2.0 Experimental, and Claude 3.5 Sonnet, using the NimStim dataset, a benchmark of 672 facial expressions (calm, angry, happy, fear, sad, neutral, surprise, and disgust) from 43 diverse human actors resulting in 2,016 model-based emotion estimates.Results. All LLM models demonstrated substantial to almost perfect agreement with ground truth labels. Happy expressions had the highest agreement, while fear had the lowest due to high misclassification as surprise. GPT-4o and Gemini had the highest accuracy, with 95% CI lower bound exceeding 0.80, while Claude performed more poorly. There were no significant differences in accuracy as a function of actor sex or race. GPT-4o and Gemini reached human performance levels for overall facial emotion recognition with GPT-4o surpassing human performance for calm/neutral and surprise recognition, while Gemini surpassed human performance for surprise recognition.Conclusion. As GenAI models increasingly mediate HCI and expand into healthcare, evaluating LLM’s socioemotional comprehension is crucial for healthcare applications. This study found that LLMs perform strongly relative to ground truth and comparably to human judges in recognizing prototypical facial expressions, with GPT-4o and Gemini models showing especially strong performance. This study lays the groundwork for evaluating the socioemotional capabilities of LLMs and highlights the need to address existing gaps for safe applications in future HCI and clinical settings.

BACKGROUND Health care chatbots are rapidly proliferating, while generative artificial intelligence (AI) outpaces existing evaluation standards. OBJECTIVE We aimed to develop a structured, stakeholder-informed framework to standardize evaluation of health care chatbots. METHODS PRISMA (Preferred Reporting Items for Systematic reviews and Meta-Analyses)–guided searches across multiple databases identified 266 records; 152 were screened, 21 full texts were assessed, and 11 frameworks were included. We extracted 356 questions (refined to 271 by deduplication and relevance review), mapped items to Coalition for Health AI constructs, and organized them with iterative input from clinicians, patients, developers, epidemiologists, and policymakers. RESULTS We developed the Health Care AI Chatbot Evaluation Framework (HAICEF), a hierarchical framework with 3 priority domains (safety, privacy, and fairness; trustworthiness and usefulness; and design and operational effectiveness) and 18 second-level and 60 third-level constructs covering 271 questions. Emphasis includes data provenance and harm control; Health Insurance Portability and Accountability Act/General Data Protection Regulation–aligned privacy and security; bias management; and reliability, transparency, and workflow integration. Question distribution across domains is as follows: design and operational effectiveness, 40%; trustworthiness and usefulness, 39%; and safety, privacy and fairness, 21%. The framework accommodates both patient-facing and back-office use cases. CONCLUSIONS HAICEF provides an adaptable scaffold for standardized evaluation and responsible implementation of health care chatbots. Planned next steps include prospective validation across settings and a Delphi consensus to extend accountability and accessibility assurances.

Background: Photoplethysmography (PPG) is a technology routinely used in clinical practice to assess blood oxygenation (SpO2) and pulse rate (PR). Skin pigmentation may influence accuracy, leading to health outcomes disparities. Objective: This systematic review and meta-analysis primarily aimed to evaluate the accuracy of PPG-derived SpO2 and PR by skin pigmentation. Secondarily, we aimed to evaluate statistical biases and the clinical relevance of PPG-derived SpO2 and PR according to skin pigmentation. Methods: We identified 23 pulse oximetry studies (n=59,684; 197,353 paired SpO2-arterial blood observations) and 4 wearable PR studies (n=176; 140,771 paired PPG-electrocardiography observations). We evaluated accuracy according to skin pigmentation group by comparing SpO2 accuracy root-mean-square values to the regulatory threshold of 3% and PR 95% limits of agreement values to +5 or –5 beats per minute (bpm), according to the standards of the American National Standards Institute, Association for the Advancement of Medical Instrumentation, and the International Electrotechnical Commission. We evaluated biases and clinical relevance using mean bias and 95% CI. Results: For SpO2, accuracy root-mean-square values were 3.96%, 4.71%, and 4.15%, and pooled mean biases were 0.70% (95% CI 0.17%-1.22%), 0.27% (95% CI –0.64% to 1.19%), and 1.27% (95% CI 0.58%-1.95%) for light, medium, and dark pigmentation, respectively. For PR, 95% limits of agreement values were from –16.02 to 13.54, from –18.62 to 16.84, and from –33.69 to 32.54, and pooled mean biases were –1.24 (95% CI –5.31 to 2.83) bpm, –0.89 (95% CI –3.70 to 1.93) bpm, and –0.57 (95% CI –9.44 to 8.29) bpm for light, medium, and dark pigmentation, respectively. Conclusions: SpO2 and PR measurements may be inaccurate across all skin pigmentation groups, breaching U.S. Food and Drug Administration guidance and industry standard thresholds. Pulse oximeters significantly overestimate SpO2 for both light and dark skin pigmentation, but this overestimation may not be clinically relevant. PRs obtained from wearables exhibit no statistically or clinically significant bias based on skin pigmentation.

Introduction: This study evaluated the performance of a wrist-worn wearable, Verily Study Watch (VSW), in detecting key sleep measures against polysomnography (PSG). Methods: We collected data from 41 adults without obstructive sleep apnea or insomnia during a single overnight laboratory visit. We evaluated epoch-by-epoch performance for sleep versus wake classification, sleep stage classification and duration, total sleep time (TST), wake after sleep onset (WASO), sleep onset latency (SOL), sleep efficiency (SE), and number of awakenings (NAWK). Performance metrics included sensitivity, specificity, Cohen’s kappa, and Bland-Altman analyses. Results: Sensitivity and specificity (95% CIs) of sleep versus wake classification were 0.97 (0.96, 0.98) and 0.70 (0.66, 0.74), respectively. Cohen’s kappa (95% CI) for 4-class stage detection was 0.64 (0.18, 0.82). Most VSW sleep measures had proportional bias. The mean bias values (95% CI) were 14.0 minutes (5.55, 23.20) for TST, - 13.1 minutes (-21.33, -6.21) for WASO, 2.97% (1.25, 4.84) for SE, -1.34 minutes (-7.29, 4.81) for SOL, 1.91 minutes (-8.28, 11.98) for light sleep duration, 5.24 minutes (-3.35, 14.13) for deep sleep duration, and 6.39 minutes (-0.68, 13.18) for REM sleep duration. Mean and median NAWK count differences (95% CI) were 0.05 (-0.42, 0.53) and 0.0 (0.0, 0.0), respectively. Discussion: Results support applying the VSW to track overnight sleep measures in free-living settings. Registered at clinicaltrials.gov ( NCT05276362 ).