Developing an individualized treatment rule for Veterans with major depressive disorder using electronic health records.

Efforts to develop an individualized treatment rule (ITR) to optimize major depressive disorder (MDD) treatment with antidepressant medication (ADM), psychotherapy, or combined ADM-psychotherapy have been hampered by small samples, small predictor sets, and suboptimal analysis methods. Analyses of large administrative databases designed to approximate experiments followed iteratively by pragmatic trials hold promise for resolving these problems. The current report presents a proof-of-concept study using electronic health records (EHR) of n = 43,470 outpatients beginning MDD treatment in Veterans Health Administration Primary Care Mental Health Integration (PC-MHI) clinics, which offer access not only to ADMs but also psychotherapy and combined ADM-psychotherapy. EHR and geospatial databases were used to generate an extensive baseline predictor set (5,865 variables). The outcome was a composite measure of at least one serious negative event (suicide attempt, psychiatric emergency department visit, psychiatric hospitalization, suicide death) over the next 12 months. Best-practices methods were used to adjust for nonrandom treatment assignment and to estimate a preliminary ITR in a 70% training sample and to evaluate the ITR in the 30% test sample. Statistically significant aggregate variation was found in overall probability of the outcome related to baseline predictors (AU-ROC = 0.68, S.E. = 0.01), with test sample outcome prevalence of 32.6% among the 5% of patients having highest predicted risk compared to 7.1% in the remainder of the test sample. The ITR found that psychotherapy-only was the optimal treatment for 56.0% of patients (roughly 20% lower risk of the outcome than if receiving one of the other treatments) and that treatment type was unrelated to outcome risk among other patients. Change in aggregate treatment costs of implementing this ITR would be negligible, as 16.1% fewer patients would be prescribed ADMs and 2.9% more would receive psychotherapy. A pragmatic trial would be needed to confirm the accuracy of the ITR.

Development of a model to predict psychotherapy response for depression among Veterans.

BACKGROUND: Fewer than half of patients with major depressive disorder (MDD) respond to psychotherapy. Pre-emptively informing patients of their likelihood of responding could be useful as part of a patient-centered treatment decision-support plan. METHODS: This prospective observational study examined a national sample of 807 patients beginning psychotherapy for MDD at the Veterans Health Administration. Patients completed a self-report survey at baseline and 3-months follow-up (data collected 2018-2020). We developed a machine learning (ML) model to predict psychotherapy response at 3 months using baseline survey, administrative, and geospatial variables in a 70% training sample. Model performance was then evaluated in the 30% test sample. RESULTS: 32.0% of patients responded to treatment after 3 months. The best ML model had an AUC (SE) of 0.652 (0.038) in the test sample. Among the one-third of patients ranked by the model as most likely to respond, 50.0% in the test sample responded to psychotherapy. In comparison, among the remaining two-thirds of patients, <25% responded to psychotherapy. The model selected 43 predictors, of which nearly all were self-report variables. CONCLUSIONS: Patients with MDD could pre-emptively be informed of their likelihood of responding to psychotherapy using a prediction tool based on self-report data. This tool could meaningfully help patients and providers in shared decision-making, although parallel information about the likelihood of responding to alternative treatments would be needed to inform decision-making across multiple treatments.

A practical risk calculator for suicidal behavior among transitioning U.S. Army soldiers: results from the Study to Assess Risk and Resilience in Servicemembers-Longitudinal Study (STARRS-LS).

BACKGROUND: Risk of suicide-related behaviors is elevated among military personnel transitioning to civilian life. An earlier report showed that high-risk U.S. Army soldiers could be identified shortly before this transition with a machine learning model that included predictors from administrative systems, self-report surveys, and geospatial data. Based on this result, a Veterans Affairs and Army initiative was launched to evaluate a suicide-prevention intervention for high-risk transitioning soldiers. To make targeting practical, though, a streamlined model and risk calculator were needed that used only a short series of self-report survey questions. METHODS: We revised the original model in a sample of n = 8335 observations from the Study to Assess Risk and Resilience in Servicemembers-Longitudinal Study (STARRS-LS) who participated in one of three Army STARRS 2011-2014 baseline surveys while in service and in one or more subsequent panel surveys (LS1: 2016-2018, LS2: 2018-2019) after leaving service. We trained ensemble machine learning models with constrained numbers of item-level survey predictors in a 70% training sample. The outcome was self-reported post-transition suicide attempts (SA). The models were validated in the 30% test sample. RESULTS: Twelve-month post-transition SA prevalence was 1.0% (s.e. = 0.1). The best constrained model, with only 17 predictors, had a test sample ROC-AUC of 0.85 (s.e. = 0.03). The 10-30% of respondents with the highest predicted risk included 44.9-92.5% of 12-month SAs. CONCLUSIONS: An accurate SA risk calculator based on a short self-report survey can target transitioning soldiers shortly before leaving service for intervention to prevent post-transition SA.

Development of a model to predict antidepressant treatment response for depression among Veterans.

BACKGROUND: Only a limited number of patients with major depressive disorder (MDD) respond to a first course of antidepressant medication (ADM). We investigated the feasibility of creating a baseline model to determine which of these would be among patients beginning ADM treatment in the US Veterans Health Administration (VHA). METHODS: A 2018-2020 national sample of n = 660 VHA patients receiving ADM treatment for MDD completed an extensive baseline self-report assessment near the beginning of treatment and a 3-month self-report follow-up assessment. Using baseline self-report data along with administrative and geospatial data, an ensemble machine learning method was used to develop a model for 3-month treatment response defined by the Quick Inventory of Depression Symptomatology Self-Report and a modified Sheehan Disability Scale. The model was developed in a 70% training sample and tested in the remaining 30% test sample. RESULTS: In total, 35.7% of patients responded to treatment. The prediction model had an area under the ROC curve (s.e.) of 0.66 (0.04) in the test sample. A strong gradient in probability (s.e.) of treatment response was found across three subsamples of the test sample using training sample thresholds for high [45.6% (5.5)], intermediate [34.5% (7.6)], and low [11.1% (4.9)] probabilities of response. Baseline symptom severity, comorbidity, treatment characteristics (expectations, history, and aspects of current treatment), and protective/resilience factors were the most important predictors. CONCLUSIONS: Although these results are promising, parallel models to predict response to alternative treatments based on data collected before initiating treatment would be needed for such models to help guide treatment selection.

Predicting suicide attempts among U.S. Army soldiers after leaving active duty using information available before leaving active duty: results from the Study to Assess Risk and Resilience in Servicemembers-Longitudinal Study (STARRS-LS).

Suicide risk is elevated among military service members who recently transitioned to civilian life. Identifying high-risk service members before this transition could facilitate provision of targeted preventive interventions. We investigated the feasibility of doing this by attempting to develop a prediction model for self-reported suicide attempts (SAs) after leaving or being released from active duty in the Study to Assess Risk and Resilience in Servicemembers-Longitudinal Study (STARRS-LS). This study included two self-report panel surveys (LS1: 2016-2018, LS2: 2018-2019) administered to respondents who previously participated while on active duty in one of three Army STARRS 2011-2014 baseline self-report surveys. We focus on respondents who left active duty >12 months before their LS survey (n = 8899). An ensemble machine learning model using predictors available prior to leaving active duty was developed in a 70% training sample and validated in a 30% test sample. The 12-month self-reported SA prevalence (SE) was 1.0% (0.1). Test sample AUC (SE) was 0.74 (0.06). The 15% of respondents with highest predicted risk included nearly two-thirds of 12-month SAs and over 80% of medically serious 12-month SAs. These results show that it is possible to identify soldiers at high post-transition self-report SA risk before the transition. Future model development is needed to examine prediction of SAs assessed by administrative data and using surveys administered closer to the time of leaving active duty.

The stories about racism and health: the development of a framework for racism narratives in medical literature using a computational grounded theory approach.

INTRODUCTION: The scientific study of racism as a root cause of health inequities has been hampered by the policies and practices of medical journals. Monitoring the discourse around racism and health inequities (i.e., racism narratives) in scientific publications is a critical aspect of understanding, confronting, and ultimately dismantling racism in medicine. A conceptual framework and multi-level construct is needed to evaluate the changes in the prevalence and composition of racism over time and across journals. OBJECTIVE: To develop a framework for classifying racism narratives in scientific medical journals. METHODS: We constructed an initial set of racism narratives based on an exploratory literature search. Using a computational grounded theory approach, we analyzed a targeted sample of 31 articles in four top medical journals which mentioned the word 'racism'. We compiled and evaluated 80 excerpts of text that illustrate racism narratives. Two coders grouped and ordered the excerpts, iteratively revising and refining racism narratives. RESULTS: We developed a qualitative framework of racism narratives, ordered on an anti-racism spectrum from impeding anti-racism to strong anti-racism, consisting of 4 broad categories and 12 granular modalities for classifying racism narratives. The broad narratives were "dismissal," "person-level," "societal," and "actionable." Granular modalities further specified how race-related health differences were related to racism (e.g., natural, aberrant, or structurally modifiable). We curated a "reference set" of example sentences to empirically ground each label. CONCLUSION: We demonstrated racism narratives of dismissal, person-level, societal, and actionable explanations within influential medical articles. Our framework can help clinicians, researchers, and educators gain insight into which narratives have been used to describe the causes of racial and ethnic health inequities, and to evaluate medical literature more critically. This work is a first step towards monitoring racism narratives over time, which can more clearly expose the limits of how the medical community has come to understand the root causes of health inequities. This is a fundamental aspect of medicine's long-term trajectory towards racial justice and health equity.

Scalable surveillance of e-cigarette products on Instagram and TikTok using computer vision.

INTRODUCTION: Instagram and TikTok, video-based social media platforms popular among adolescents, contain tobacco-related content despite the platforms' policies prohibiting substance-related posts. Prior research identified themes in e-cigarette-related social media posts using qualitative or text-based machine learning methods. We developed an image-based computer vision model to identify e-cigarette products in social media images and videos. METHODS: We created a dataset of 6,999 Instagram images labeled for 8 object classes: mod or pod devices, e-juice containers, packaging boxes, nicotine warning labels, e-juice flavors, e-cigarette brand names, and smoke clouds. We trained a DyHead object detection model using a Swin-Large backbone, evaluated the model's performance on 20 Instagram and TikTok videos, and applied the model to 14,072 e-cigarette-related promotional TikTok videos (2019-2022; 10,276,485 frames). RESULTS: The model achieved the following mean average precision scores on the image test set: e-juice container: 0.89; pod device: 0.67; mod device: 0.54; packaging box: 0.84; nicotine warning label: 0.86; e-cigarette brand name: 0.71; e-juice flavor name: 0.89; and smoke cloud: 0.46. The largest number of TikTok videos - 9,091 (65%) - contained smoke clouds, followed by mod and pod devices detected in 6,667 (47%) and 5,949 (42%) videos respectively. Prevalence of nicotine warning labels was the lowest, detected in 980 videos (7%). CONCLUSIONS: Deep learning-based object detection technology enables automated analysis of visual posts on social media. Our computer vision model can detect the presence of e-cigarettes products in images and videos, providing valuable surveillance data for tobacco regulatory science. IMPLICATIONS: Prior research identified themes in e-cigarette-related social media posts using qualitative or text-based machine learning methods. We developed an image-based computer vision model to identify e-cigarette products in social media images and videos.We trained a DyHead object detection model using a Swin-Large backbone, evaluated the model's performance on 20 Instagram and TikTok videos featuring at least two e-cigarette objects, and applied the model to 14,072 e-cigarette-related promotional TikTok videos (2019-2022; 10,276,485 frames).The deep learning model can be used for automated, scalable surveillance of image- and video-based e-cigarette-related promotional content on social media, providing valuable data for tobacco regulatory science. Social media platforms could use computer vision to identify tobacco-related imagery and remove it promptly, which could reduce adolescents' exposure to tobacco content online.

Distinguishing Admissions Specifically for COVID-19 From Incidental SARS-CoV-2 Admissions: National Retrospective Electronic Health Record Study.

BACKGROUND: Admissions are generally classified as COVID-19 hospitalizations if the patient has a positive SARS-CoV-2 polymerase chain reaction (PCR) test. However, because 35% of SARS-CoV-2 infections are asymptomatic, patients admitted for unrelated indications with an incidentally positive test could be misclassified as a COVID-19 hospitalization. Electronic health record (EHR)-based studies have been unable to distinguish between a hospitalization specifically for COVID-19 versus an incidental SARS-CoV-2 hospitalization. Although the need to improve classification of COVID-19 versus incidental SARS-CoV-2 is well understood, the magnitude of the problems has only been characterized in small, single-center studies. Furthermore, there have been no peer-reviewed studies evaluating methods for improving classification. OBJECTIVE: The aims of this study are to, first, quantify the frequency of incidental hospitalizations over the first 15 months of the pandemic in multiple hospital systems in the United States and, second, to apply electronic phenotyping techniques to automatically improve COVID-19 hospitalization classification. METHODS: From a retrospective EHR-based cohort in 4 US health care systems in Massachusetts, Pennsylvania, and Illinois, a random sample of 1123 SARS-CoV-2 PCR-positive patients hospitalized from March 2020 to August 2021 was manually chart-reviewed and classified as "admitted with COVID-19" (incidental) versus specifically admitted for COVID-19 ("for COVID-19"). EHR-based phenotyping was used to find feature sets to filter out incidental admissions. RESULTS: EHR-based phenotyped feature sets filtered out incidental admissions, which occurred in an average of 26% of hospitalizations (although this varied widely over time, from 0% to 75%). The top site-specific feature sets had 79%-99% specificity with 62%-75% sensitivity, while the best-performing across-site feature sets had 71%-94% specificity with 69%-81% sensitivity. CONCLUSIONS: A large proportion of SARS-CoV-2 PCR-positive admissions were incidental. Straightforward EHR-based phenotypes differentiated admissions, which is important to assure accurate public health reporting and research.

Evaluation of U.S. state opioid prescribing restrictions using patient opioid consumption patterns from a single, urban, academic institution.

Background: Since 2017, states, insurers, and pharmacies have placed blanket limits on the duration and quantity of opioid prescriptions. In many states, overlapping duration and daily dose limits yield maximum prescription limits of 150-350 morphine milligram equivalents (MMEs). There is limited knowledge of how these restrictions compare with actual patient opioid consumption; while changes in prescription patterns and opioid misuse rates have been studied, these are, at best, weak proxies for actual pain control consumption. We sought to determine how patients undergoing surgery would be affected by opioid prescribing restrictions using actual patient opioid consumption data. Methods: We constructed a prospective database of post-discharge opioid consumption: patients undergoing surgery at one institution were called after discharge to collect opioid consumption data. Patients whose opioid consumption exceeded 150 and 350 MME were identified. Results: Two thousand nine hundred and seventy-one patients undergoing 54 common surgical procedures were included in our study. Twenty-one percent of patients consumed more than the 150 MME limit. Only 7% of patients consumed above the 350 MME limit. Typical (non-outlier) opioid consumption, defined as less than the 75th percentile of consumption for any given procedure, exceeded the 150 MME and 350 MME limits for 41 and 7% of procedures, respectively. Orthopedic, spinal/neurosurgical, and complex abdominal procedures most commonly exceeded these limits. Conclusions: While most patients undergoing surgery are unaffected by recent blanket prescribing limits, those undergoing a specific subset of procedures are likely to require more opioids than the restrictions permit; providers should be aware that these patients may require a refill to adequately control post-surgical pain. Real consumption data should be used to guide these restrictions and inform future interventions so the risk of worsened pain control (and its troublesome effects on opioid misuse) is minimized. Procedure-specific prescribing limits may be one approach to prevent misuse, while also optimizing post-operative pain control.

Invited Commentary: New Directions in Machine Learning Analyses of Administrative Data to Prevent Suicide-Related Behaviors.

This issue contains a thoughtful report by Gradus et al. (Am J Epidemiol. 2021;190(12):2517-2527) on a machine learning analysis of administrative variables to predict suicide attempts over 2 decades throughout Denmark. This is one of numerous recent studies that document strong concentration of risk of suicide-related behaviors among patients with high scores on machine learning models. The clear exposition of Gradus et al. provides an opportunity to review major challenges in developing, interpreting, and using such models: defining appropriate controls and time horizons, selecting comprehensive predictors, dealing with imbalanced outcomes, choosing classifiers, tuning hyperparameters, evaluating predictor variable importance, and evaluating operating characteristics. We close by calling for machine-learning research into suicide-related behaviors to move beyond merely demonstrating significant prediction-this is by now well-established-and to focus instead on using such models to target specific preventive interventions and to develop individualized treatment rules that can be used to help guide clinical decisions to address the growing problems of suicide attempts, suicide deaths, and other injuries and deaths in the same spectrum.

Evaluation of a city-wide school-located influenza vaccination program in Oakland, California, with respect to vaccination coverage, school absences, and laboratory-confirmed influenza: A matched cohort study.

BACKGROUND: It is estimated that vaccinating 50%-70% of school-aged children for influenza can produce population-wide indirect effects. We evaluated a city-wide school-located influenza vaccination (SLIV) intervention that aimed to increase influenza vaccination coverage. The intervention was implemented in ≥95 preschools and elementary schools in northern California from 2014 to 2018. Using a matched cohort design, we estimated intervention impacts on student influenza vaccination coverage, school absenteeism, and community-wide indirect effects on laboratory-confirmed influenza hospitalizations. METHODS AND FINDINGS: We used a multivariate matching algorithm to identify a nearby comparison school district with pre-intervention characteristics similar to those of the intervention school district and matched schools in each district. To measure student influenza vaccination, we conducted cross-sectional surveys of student caregivers in 22 school pairs (2017 survey, N = 6,070; 2018 survey, N = 6,507). We estimated the incidence of laboratory-confirmed influenza hospitalization from 2011 to 2018 using surveillance data from school district zip codes. We analyzed student absenteeism data from 2011 to 2018 from each district (N = 42,487,816 student-days). To account for pre-intervention differences between districts, we estimated difference-in-differences (DID) in influenza hospitalization incidence and absenteeism rates using generalized linear and log-linear models with a population offset for incidence outcomes. Prior to the SLIV intervention, the median household income was $51,849 in the intervention site and $61,596 in the comparison site. The population in each site was predominately white (41% in the intervention site, 48% in the comparison site) and/or of Hispanic or Latino ethnicity (26% in the intervention site, 33% in the comparison site). The number of students vaccinated by the SLIV intervention ranged from 7,502 to 10,106 (22%-28% of eligible students) each year. During the intervention, influenza vaccination coverage among elementary students was 53%-66% in the comparison district. Coverage was similar between the intervention and comparison districts in influenza seasons 2014-2015 and 2015-2016 and was significantly higher in the intervention site in seasons 2016-2017 (7%; 95% CI 4, 11; p < 0.001) and 2017-2018 (11%; 95% CI 7, 15; p < 0.001). During seasons when vaccination coverage was higher among intervention schools and the vaccine was moderately effective, there was evidence of statistically significant indirect effects: The DID in the incidence of influenza hospitalization per 100,000 in the intervention versus comparison site was -17 (95% CI -30, -4; p = 0.008) in 2016-2017 and -37 (95% CI -54, -19; p < 0.001) in 2017-2018 among non-elementary-school-aged individuals and -73 (95% CI -147, 1; p = 0.054) in 2016-2017 and -160 (95% CI -267, -53; p = 0.004) in 2017-2018 among adults 65 years or older. The DID in illness-related school absences per 100 school days during the influenza season was -0.63 (95% CI -1.14, -0.13; p = 0.014) in 2016-2017 and -0.80 (95% CI -1.28, -0.31; p = 0.001) in 2017-2018. Limitations of this study include the use of an observational design, which may be subject to unmeasured confounding, and caregiver-reported vaccination status, which is subject to poor recall and low response rates. CONCLUSIONS: A city-wide SLIV intervention in a large, diverse urban population was associated with a decrease in the incidence of laboratory-confirmed influenza hospitalization in all age groups and a decrease in illness-specific school absence rate among students in 2016-2017 and 2017-2018, seasons when the vaccine was moderately effective, suggesting that the intervention produced indirect effects. Our findings suggest that in populations with moderately high background levels of influenza vaccination coverage, SLIV programs are associated with further increases in coverage and reduced influenza across the community.

Early immune stimulation and childhood acute lymphoblastic leukemia in Costa Rica: A comparison of statistical approaches.

BACKGROUND: Although epidemiologic studies suggest that early immune stimulation is protective against childhood leukemia, evidence for this relationship is equivocal for Hispanic children, who are disproportionately affected by this disease. The complex biological processes underlying immune stimulation and leukemogenesis may benefit from novel statistical approaches that account for mixed exposures and their nonlinear interactions. In this study, we utilized targeted machine learning and traditional statistical methods to investigate the association of multiple measures of early immune stimulation with acute lymphoblastic leukemia (ALL) in Costa Rican children. MATERIALS AND METHODS: We used data from a population-based case-control study conducted in Costa Rica (2001-2003). Cases of ALL (n = 240) were diagnosed in 1995-2000 (age >1 year and <15 years at diagnosis) and were identified through the National Cancer Registry and National Children's Hospital. Population controls (n = 578) were frequency-matched to cases by birth year and drawn from the National Birth Registry. Data on surrogate measures of early immune stimulation were collected through in-home interviews. We fitted multivariable models, utilizing targeted causal inference (varimpact), unconditional logistic regression, and latent class analysis (LCA). RESULTS: In varimpact analysis, contact with any pet [risk difference (RD) = -0.17, 95% CI: -0.25, -0.10)] or any farm animal (RD = -0.07, 95% CI: -0.13, 0.00) and allergies (RD = -0.08, 95% CI: -0.17, 0.01) were associated with a reduced risk of ALL, whereas experiencing a fever longer than one week was associated with an increased risk (RD = 0.23, 95% CI: 0.12, 0.33). In unconditional logistic regression models, contact with any pet or farm animal and a complete vaccination scheme were inversely associated with odds of ALL (OR = 0.44, 95% CI: 0.31, 0.62; OR = 0.66, 95% CI: 0.49, 0.90; OR = 0.45, 95% CI: 0.24, 0.83; respectively), whereas experiencing a fever longer than one week was positively associated with ALL (OR = 2.44, 95% CI: 1.61, 3.70). Two-class and three-class LCA revealed a group with elevated risk for ALL whose exposure profile was mainly characterized by reduced exposure to pets and farm animals. CONCLUSIONS: Using distinct statistical approaches, we observed that exposure to pets and farm animals was inversely associated with ALL risk, whereas having a fever longer than one week (a putative proxy of severe infection) was associated with an increased risk. For multifactorial diseases such as childhood leukemia, we recommend estimating the joint effects of multiple exposures by applying diverse statistical methods and interpreting their results together. Overall, we found support for the hypothesis that early immune stimulation offers protection against childhood ALL.

60-day major adverse cardiac events in emergency department patients with non-low modified HEART scores.

BACKGROUND: A low (0-3) History, Electrocardiogram, Age, Risk factors and Troponin (HEART) score reliably identifies ED chest pain patients who are low risk for near-term major adverse cardiac events (MACE). To optimize sensitivity, many clinicians employ a modified HEART score by repeating troponin measurements and excluding patients with abnormal troponin values or ischemic electrocardiograms (ECGs). The residual MACE risk among patients with otherwise non-low (≥4) modified HEART scores is thus likely much lower than with non-low original HEART scores. OBJECTIVE: To explore residual 60-day MACE risks among patients with non-low modified HEART scores. METHODS: Secondary analysis of a retrospective cohort of ED patients presenting with chest pain to an integrated healthcare system between 2013 and 2015. Patients with serial troponin measurements within 6 h of ED arrival were considered for inclusion. Exclusions included an ischemic ECG, troponin values above the 99th percentile or a lack of continuous health plan coverage through the 60-day follow-up period. MACE was defined as a composite of myocardial infarction, cardiac arrest, cardiogenic shock or death. RESULTS: There were 22,976 study eligible patients encounters, 13,521 (59%) of which had non-low (≥4) modified HEART scores. The observed 60-day MACE risk among non-low HEART score patients was 2.0% (95% CI 1.8-2.3). When including all coronary revascularizations (MACE-R), the risk was 4.4% (95% CI 4.1-4.4). CONCLUSION: Risk of near-term MACE among patients with non-low modified HEART scores (excluding those with abnormal troponin or ischemic ECGs) appears to be much lower than in the original HEART score validation studies.

Patterns and correlates of mental healthcare utilization during the COVID-19 pandemic among individuals with pre-existing mental disorder.

How did mental healthcare utilization change during the COVID-19 pandemic period among individuals with pre-existing mental disorder? Understanding utilization patterns of these at-risk individuals and identifying those most likely to exhibit increased utilization could improve patient stratification and efficient delivery of mental health services. This study leveraged large-scale electronic health record (EHR) data to describe mental healthcare utilization patterns among individuals with pre-existing mental disorder before and during the COVID-19 pandemic and identify correlates of high mental healthcare utilization. Using EHR data from a large healthcare system in Massachusetts, we identified three "pre-existing mental disorder" groups (PMD) based on having a documented mental disorder diagnosis within the 6 months prior to the March 2020 lockdown, related to: (1) stress-related disorders (e.g., depression, anxiety) (N = 115,849), (2) serious mental illness (e.g., schizophrenia, bipolar disorders) (N = 11,530), or (3) compulsive behavior disorders (e.g., eating disorder, OCD) (N = 5,893). We also identified a "historical comparison" group (HC) for each PMD (N = 113,604, 11,758, and 5,387, respectively) from the previous year (2019). We assessed the monthly number of mental healthcare visits from March 13 to December 31 for PMDs in 2020 and HCs in 2019. Phenome-wide association analyses (PheWAS) were used to identify clinical correlates of high mental healthcare utilization. We found the overall number of mental healthcare visits per patient during the pandemic period in 2020 was 10-12% higher than in 2019. The majority of increased visits was driven by a subset of high mental healthcare utilizers (top decile). PheWAS results indicated that correlates of high utilization (prior mental disorders, chronic pain, insomnia, viral hepatitis C, etc.) were largely similar before and during the pandemic, though several conditions (e.g., back pain) were associated with high utilization only during the pandemic. Limitations included that we were not able to examine other risk factors previously shown to influence mental health during the pandemic (e.g., social support, discrimination) due to lack of social determinants of health information in EHR data. Mental healthcare utilization among patients with pre-existing mental disorder increased overall during the pandemic, likely due to expanded access to telemedicine. Given that clinical correlates of high mental healthcare utilization in a major hospital system were largely similar before and during the COVID-19 pandemic, resource stratification based on known risk factor profiles may aid hospitals in responding to heightened mental healthcare needs during a pandemic.

Analyzing Surgical Technique in Diverse Open Surgical Videos With Multitask Machine Learning.

Objective: To overcome limitations of open surgery artificial intelligence (AI) models by curating the largest collection of annotated videos and to leverage this AI-ready data set to develop a generalizable multitask AI model capable of real-time understanding of clinically significant surgical behaviors in prospectively collected real-world surgical videos. Design, Setting, and Participants: The study team programmatically queried open surgery procedures on YouTube and manually annotated selected videos to create the AI-ready data set used to train a multitask AI model for 2 proof-of-concept studies, one generating surgical signatures that define the patterns of a given procedure and the other identifying kinematics of hand motion that correlate with surgeon skill level and experience. The Annotated Videos of Open Surgery (AVOS) data set includes 1997 videos from 23 open-surgical procedure types uploaded to YouTube from 50 countries over the last 15 years. Prospectively recorded surgical videos were collected from a single tertiary care academic medical center. Deidentified videos were recorded of surgeons performing open surgical procedures and analyzed for correlation with surgical training. Exposures: The multitask AI model was trained on the AI-ready video data set and then retrospectively applied to the prospectively collected video data set. Main Outcomes and Measures: Analysis of open surgical videos in near real-time, performance on AI-ready and prospectively collected videos, and quantification of surgeon skill. Results: Using the AI-ready data set, the study team developed a multitask AI model capable of real-time understanding of surgical behaviors-the building blocks of procedural flow and surgeon skill-across space and time. Through principal component analysis, a single compound skill feature was identified, composed of a linear combination of kinematic hand attributes. This feature was a significant discriminator between experienced surgeons and surgical trainees across 101 prospectively collected surgical videos of 14 operators. For each unit increase in the compound feature value, the odds of the operator being an experienced surgeon were 3.6 times higher (95% CI, 1.67-7.62; P = .001). Conclusions and Relevance: In this observational study, the AVOS-trained model was applied to analyze prospectively collected open surgical videos and identify kinematic descriptors of surgical skill related to efficiency of hand motion. The ability to provide AI-deduced insights into surgical structure and skill is valuable in optimizing surgical skill acquisition and ultimately improving surgical care.

Toward personalized care for insomnia in the US Army: a machine learning model to predict response to cognitive behavioral therapy for insomnia.

STUDY OBJECTIVES: The standard of care for military personnel with insomnia is cognitive behavioral therapy for insomnia (CBT-I). However, only a minority seeking insomnia treatment receive CBT-I, and little reliable guidance exists to identify those most likely to respond. As a step toward personalized care, we present results of a machine learning (ML) model to predict CBT-I response. METHODS: Administrative data were examined for n = 1,449 nondeployed US Army soldiers treated for insomnia with CBT-I who had moderate-severe baseline Insomnia Severity Index (ISI) scores and completed 1 or more follow-up ISIs 6-12 weeks after baseline. An ensemble ML model was developed in a 70% training sample to predict clinically significant ISI improvement (reduction of at least 2 standard deviations on the baseline ISI distribution). Predictors included a wide range of military administrative and baseline clinical variables. Model accuracy was evaluated in the remaining 30% test sample. RESULTS: 19.8% of patients had clinically significant ISI improvement. Model area under the receiver operating characteristic curve (standard error) was 0.60 (0.03). The 20% of test-sample patients with the highest probabilities of improvement were twice as likely to have clinically significant improvement compared with the remaining 80% (36.5% vs 15.7%; χ21 = 9.2, P = .002). Nearly 85% of prediction accuracy was due to 10 variables, the most important of which were baseline insomnia severity and baseline suicidal ideation. CONCLUSIONS: Pending replication, the model could be used as part of a patient-centered decision-making process for insomnia treatment. Parallel models will be needed for alternative treatments before such a system is of optimal value. CITATION: Gabbay FH, Wynn GH, Georg MW, et al. Toward personalized care for insomnia in the US Army: a machine learning model to predict response to cognitive behavioral therapy for insomnia. J Clin Sleep Med. 2024;20(6):921-931.

Comparative effectiveness of remote digital gamified and group CBT skills training interventions for anxiety and depression among college students: Results of a three-arm randomised controlled trial.

Digital interventions can enhance access to healthcare in under-resourced settings. However, guided digital interventions may be costly for low- and middle-income countries, despite their effectiveness. In this randomised control trial, we evaluated the effectiveness of two digital interventions designed to address this issue: (1) a Cognitive Behavioral Therapy Skills Training (CST) intervention that increased scalability by using remote online group administration; and (2) the SuperBetter gamified self-guided CBT skills training app, which uses other participants rather than paid staff as guides. The study was implemented among anxious and/or depressed South African undergraduates (n = 371) randomised with equal allocation to Remote Group CST, SuperBetter, or a MoodFlow mood monitoring control. Symptoms were assessed with the Generalized Anxiety Disorder-7 (GAD-7) and the Patient Health Questionnaire-9 (PHQ-9). Intention-to-treat analysis found effect sizes at the high end of prior digital intervention trials, including significantly higher adjusted risk differences (ARD; primary outcome) in joint anxiety/depression remission at 3-months and 6-months for Remote Group CST (ARD = 23.3-18.9%, p = 0.001-0.035) and SuperBetter (ARD = 12.7-22.2%, p = 0.047-0.006) than MoodFlow and mean combined PHQ-9/GAD-7 scores (secondary outcome) significantly lower for Remote Group CST and SuperBetter than MoodFlow. These results illustrate how innovative delivery methods can increase the scalability of standard one-on-one guided digital interventions. PREREGISTRATION INTERNATIONAL STANDARD RANDOMISED CONTROLLED TRIAL NUMBER (ISRTCN) SUBMISSION #: 47,089,643.

Estimated Average Treatment Effect of Psychiatric Hospitalization in Patients With Suicidal Behaviors: A Precision Treatment Analysis.

Importance: Psychiatric hospitalization is the standard of care for patients presenting to an emergency department (ED) or urgent care (UC) with high suicide risk. However, the effect of hospitalization in reducing subsequent suicidal behaviors is poorly understood and likely heterogeneous. Objectives: To estimate the association of psychiatric hospitalization with subsequent suicidal behaviors using observational data and develop a preliminary predictive analytics individualized treatment rule accounting for heterogeneity in this association across patients. Design, Setting, and Participants: A machine learning analysis of retrospective data was conducted. All veterans presenting with suicidal ideation (SI) or suicide attempt (SA) from January 1, 2010, to December 31, 2015, were included. Data were analyzed from September 1, 2022, to March 10, 2023. Subgroups were defined by primary psychiatric diagnosis (nonaffective psychosis, bipolar disorder, major depressive disorder, and other) and suicidality (SI only, SA in past 2-7 days, and SA in past day). Models were trained in 70.0% of the training samples and tested in the remaining 30.0%. Exposures: Psychiatric hospitalization vs nonhospitalization. Main Outcomes and Measures: Fatal and nonfatal SAs within 12 months of ED/UC visits were identified in administrative records and the National Death Index. Baseline covariates were drawn from electronic health records and geospatial databases. Results: Of 196 610 visits (90.3% men; median [IQR] age, 53 [41-59] years), 71.5% resulted in hospitalization. The 12-month SA risk was 11.9% with hospitalization and 12.0% with nonhospitalization (difference, -0.1%; 95% CI, -0.4% to 0.2%). In patients with SI only or SA in the past 2 to 7 days, most hospitalization was not associated with subsequent SAs. For patients with SA in the past day, hospitalization was associated with risk reductions ranging from -6.9% to -9.6% across diagnoses. Accounting for heterogeneity, hospitalization was associated with reduced risk of subsequent SAs in 28.1% of the patients and increased risk in 24.0%. An individualized treatment rule based on these associations may reduce SAs by 16.0% and hospitalizations by 13.0% compared with current rates. Conclusions and Relevance: The findings of this study suggest that psychiatric hospitalization is associated with reduced average SA risk in the immediate aftermath of an SA but not after other recent SAs or SI only. Substantial heterogeneity exists in these associations across patients. An individualized treatment rule accounting for this heterogeneity could both reduce SAs and avert hospitalizations.

Predicting Homelessness Among Transitioning U.S. Army Soldiers.

INTRODUCTION: This study develops a practical method to triage Army transitioning service members (TSMs) at highest risk of homelessness to target a preventive intervention. METHODS: The sample included 4,790 soldiers from the Study to Assess Risk and Resilience in Servicemembers-Longitudinal Study (STARRS-LS) who participated in 1 of 3 Army STARRS 2011-2014 baseline surveys followed by the third wave of the STARRS-LS online panel surveys (2020-2022). Two machine learning models were trained: a Stage-1 model that used administrative predictors and geospatial data available for all TSMs at discharge to identify high-risk TSMs for initial outreach; and a Stage-2 model estimated in the high-risk subsample that used self-reported survey data to help determine highest risk based on additional information collected from high-risk TSMs once they are contacted. The outcome in both models was homelessness within 12 months after leaving active service. RESULTS: Twelve-month prevalence of post-transition homelessness was 5.0% (SE=0.5). The Stage-1 model identified 30% of high-risk TSMs who accounted for 52% of homelessness. The Stage-2 model identified 10% of all TSMs (i.e., 33% of high-risk TSMs) who accounted for 35% of all homelessness (i.e., 63% of the homeless among high-risk TSMs). CONCLUSIONS: Machine learning can help target outreach and assessment of TSMs for homeless prevention interventions.