Samrachana Adhikari

Objective: Using AI algorithms can exacerbate health disparities if care or resources are allocated away from underserved populations. We evaluated an algorithm for its potential to worsen health disparities across different clinical use cases. Materials and Methods: This was a retrospective study of patients with heart failure (HF) at an academic health system using an algorithm that predicts pharmacy fill nonadherence to evidence-based HF medications. We compared prediction performance metrics (accuracy, false positive rate, false negative rate), using rate-ratios (RRs), between subgroups with and without known HF care disparities: below vs above median neighborhood-level socioeconomic status (nSES) and Black vs White race. Results were then applied to 3 hypothetical clinical use cases. Results: Among 34 697 patients (13% Black, 10% Hispanic, 65% White), algorithm accuracy was similar across nSES and racial subgroups. The algorithm assigned more false positives for medication nonadherence among low vs high nSES (RR [95%CI] 1.50 [1.44-1.56]) and Black vs White (2.05 [1.92-2.19]) subgroups. The algorithm also assigned fewer false negatives (0.63 [0.59-0.67]) to Black vs White subgroups. When applied to 3 hypothetical use cases, worsening of existing disparities was pertinent for clinical applications where false positives could be particularly harmful (e.g, if predictions of nonadherence prompted lower treatment priority). Discussion: Although accuracy was similar across demographic groups, differences in false positive and false negative rates revealed that the same prediction may worsen disparities in some use cases, but not others. Conclusion: Evaluation of predictions in the context of clinical use is essential to avoid unintentionally worsening inequities.

Background: Ketamine is increasingly prescribed in an off‐label manner to treat psychiatric disorders, raising concerns about direct‐to‐consumer advertising and the proliferation of clinics offering ketamine for at‐home use without direct medical supervision. Objectives: We aimed to identify and characterize clinics advertising ketamine for psychiatric conditions online in the New York metropolitan area (New York, New Jersey, and Connecticut), with attention to advertising suggesting ketamine is prescribed for at‐home use. Methods: In 2025, systematic web searches were conducted to identify clinics advertising prescription ketamine for psychiatric indications. Public‐facing website content was reviewed to describe clinic characteristics: service delivery modality, clinician credentials, routes of administration, disorders treated, and advertising practices. A generalized linear model was used to delineate correlates of clinics advertising ketamine for at‐home use. Results: 233 clinics were located; 36.5% prescribed ketamine for at‐home use. 51.5% listed a medical doctor as part of their team and 42.9% advertised oral ketamine. Depression was the most commonly listed disorder treated (94.0%) and 21.9% advertised ketamine to treat substance use disorder. In the multivariable model, advertising ketamine for at‐home use was more common among clinics advertising oral ketamine (aPR = 4.10, 95% CI: 2.20–7.60) and less common among clinics listing a medical doctor (aPR = 0.54, 95% CI: 0.30–0.99). Conclusions: Over a third of clinics advertised ketamine for at‐home use. A limitation is that we only focused on public‐facing websites. Advertising practices and clinician representation suggest clinics may be advertising in a more consumer‐oriented manner, underscoring the need for monitoring and clearer guidance to mitigate potential safety risks.

Randomized controlled trials (RCTs) often include subgroup analyses to assess whether treatment effects vary across prespecified patient populations. However, these analyses frequently suffer from small sample sizes, which limit the power to detect heterogeneous effects. Power can be improved by leveraging predictors of the outcome—that is, through covariate adjustment—as well as by borrowing external data from similar RCTs or observational studies. The benefits of covariate adjustment may be limited when the trial sample is small. Borrowing external data can increase the effective sample size and improve power, but it introduces two key challenges: (i) integrating data across sources can lead to model misspecification, and (ii) practical violations of the positivity assumption—where the probability of receiving the target treatment is near zero for some covariate profiles in the external data—can lead to extreme inverse‐probability weights and unstable inferences, ultimately negating potential power gains. To account for these shortcomings, we present an approach to improving power in preplanned subgroup analyses of small RCTs that leverages both baseline predictors and external data. We propose de‐biased estimators that accommodate parametric, machine learning (ML), and nonparametric Bayesian methods. To address practical positivity violations (PPVs), we introduce three estimators: A covariate‐balancing approach, an automated de‐biased machine learning (DML) estimator, and a calibrated‐DML estimator. We show improved power in various simulations and offer practical recommendations for the application of the proposed methods. Finally, we apply them to evaluate the effectiveness of citalopram for negative symptoms in first‐episode schizophrenia (FES) patients across subgroups defined by duration of untreated psychosis (DUP), using data from two small RCTs.

Background: Wearable accelerometers, which continuously record physical activity metrics, are commonly used in mobile health–enabled cardiac rehabilitation (mHealth-CR). The association between adherence to accelerometer use during mHealth-CR and improvement in clinical outcomes, such as functional capacity, is understudied. The emergence of artificial intelligence (AI) technology provides novel opportunities to investigate accelerometry use patterns in relation to mHealth-CR outcomes. Objective: In this study, we sought to use an AI clustering framework to identify distinct behavioral phenotypes of adherence to accelerometer use. We then aimed to quantify the association of these adherence phenotypes with functional capacity improvements in older adults undergoing mHealth-CR. Methods: We analyzed data from the RESILIENT (Rehabilitation at Home Using Mobile Health in Older Adults After Hospitalization for Ischemic Heart Disease) trial, the largest randomized clinical study to date comparing mHealth-CR versus usual care in older adults (aged ≥65 years). Intervention arm participants were instructed to wear a Fitbit accelerometer for the 3-month study duration. Adherence to accelerometer use was quantified as overall adherence (percentage of days worn) via k-means clustering AI-derived measures and compared with changes in 6-minute walk distance (6-MWD), adjusted for demographic and clinical covariates. Results: Among 271 participants with a mean age of 71 years (SD 8), of whom 198 (73%) were male, accelerometers were worn for an average of 76 days (95% confidence limits 73,78) over 3 months. Adjusted analyses showed a weak association between days of wear and improvement in 6-MWD, with every 30 additional days associated with an 11-meter improvement (P=.08). Our k-means clustering framework identified adherence phenotypes at two resolutions: low resolution (k=2 clusters) and high resolution (k=8 clusters). The consistently high adherence cluster trended toward a 24.6-meter improvement in 6-MWD compared to the low and declining adherence clusters (n=39; 95% CI 0.7-49.9; P=.06). The 8-cluster phenotyping revealed a richer set of adherence patterns, with the consistently high adherence cluster in this analysis having a 38.5-meter (95% CI 2.2-74.7; P=.04) improvement in 6-MWD than the low adherence cluster, as well as greater average daily steps over the 3-month intervention (mean 7518, SD 3415 vs mean 4800, SD 2920 steps; P=.008). Conclusions: A time-series AI clustering framework identified a range of behavioral phenotypes representing different degrees of adherence to accelerometer use. Regression analysis identified a weak association between the higher adherence phenotype and functional capacity improvement in older adults undergoing mHealth-CR. Our AI-derived accelerometry adherence phenotypes may offer a new approach to tailor mHealth-CR regimens to individual patients, potentially leading to better outcomes in this high-risk population. Trial Registration: ClinicalTrials.gov NCT03978130; https://clinicaltrials.gov/study/NCT03978130 International Registered Report Identifier (IRRID): RR2-10.2196/32163

Background: There is no gold standard for measuring adherence to a complex medication regimen. Heart failure is a chronic disease state that requires multiple medications for optimal control, known as guideline-directed-medical therapy (GDMT), and can be used a model to explore approaches to assessing multi-medication adherence. We aimed to compare seven proportion of days covered (PDC) measures for assessing adherence to multiple GDMT medications and to evaluate their association with clinical outcomes. Methods: We conducted a large, single center, retrospective cohort study of 34,603 patients with heart failure who filled a GDMT prescription between April 2021 and July 2022. The primary outcomes were the seven PDC-based adherence measures derived from electronic health record and pharmacy data. PDC was defined as the proportion of days in which patients had possession of GDMT medications. Our secondary outcome was the combined clinical outcome of ED visits, hospitalizations, and death. Results: The seven measures provided a wide range of measured PDC adherence. Using the strictest measure, ‘each’ (>= 80% of days with each drug available) 54% of patients were considered adherent, compared with 73% measured adherence for the least restrictive measure, ‘at least one’. Variability increased with increasing number of medications across all non-average based measures. For example, using the ‘all’ measure (a more restrictive PDC measure) adherence ranged from 0.53 to 0.41 with increasing number of GDMT prescriptions. Higher PDC for each of the seven measures was associated with increased number of ED, visits, hospitalizations, and death. There was no association with the combined outcome in patients with heart failure with reduced ejection fraction. Conclusions and Relevance: There was a wide variability in adherence measures for assessing adherence to GDMT depending on the measure used. This variability has significant implications for the policy, clinical, and intervention context to which the measure is applied.

Homework is a potential contributor to treatment response in cognitive-behavioral therapy (CBT) for anxiety, but less is known regarding the importance of yoga homework for generalized anxiety disorder (GAD). This study examined the impact of homework engagement on treatment response within a randomized controlled trial (RCT) of 12 weeks of group CBT, Kundalini Yoga (KY), or stress education (SE) in a subsample of 190 adults with GAD (71% female, Mean age = 33 ± 13) who attended ≥2 sessions and submitted ≥1 homework log. Participants in CBT and KY showed greater overall homework engagement than those in SE (ps < .05). Across treatment arms, staff-rated homework compliance (p = .002, OR = 1.74), but not participant-reported days per week engaged in homework (p = .108), predicted clinical response at post-treatment ("response"). Greater staff-rated homework compliance was related to a greater response for those in CBT (p = .005, OR = 2.49) and KY (p = .049, OR = 1.66), but not SE. Greater participant-reported homework days per week was only marginally related to response to CBT (p = .054, OR = 1.71), and was not related to response to KY or SE. These findings highlight the importance of homework engagement in CBT for GAD. More research is needed to further elucidate the role of homework engagement in yoga for GAD.TRIAL REGISTRATION: clinicaltrials.gov: NCT01912287; https://clinicaltrials.gov/ct2/show/NCT01912287.

Background: Patients with heart failure (HF) often have difficulty obtaining life‐saving medications due to coverage barriers, such as prior authorizations and high out‐of‐pocket costs. To promote better coverage for high value therapies and inform policymakers about cost effectiveness, the American Heart Association/American College of Cardiology/Heart Failure Society of America added Value Statements to HF guidelines. We assessed whether these guidelines influenced Medicare drug coverage policies for 2 life‐saving, costly HF medications: angiotensin receptor neprilysin inhibitors (ARNI–guideline “high value”) and sodium glucose cotransporter‐2 inhibitors (SGLT2i–guideline “intermediate value”). Methods: We performed an observational study using Medicare prescription drug plan formulary files from April 2020 to April 2023 to separately assess for changes in coverage barriers to ARNI and SGLT2i after Value Statement publication (April 2022), and subsequent Medicare plan online update (October 2022). The primary outcome was the percentage of plans each month with any barrier to drug coverage (prior authorizations, tier ≥3 out‐of‐pocket cost‐sharing, step therapy, or no coverage). Analyses used interrupted time series and difference‐in‐differences approaches. Difference‐in‐differences analyses used direct oral anticoagulants as a control due to their comparable cost and use as ARNI and SGLT2i, but without a Value Statement. Results: Among 7396 Medicare drug plans, monthly rates of any coverage barrier ranged from 94.3% to 97.4% for ARNI and 93.2% to 96.6% for SGLT2i. Most barriers were due to tier ≥3 out‐of‐pocket cost‐sharing requirements (ARNI: 94.3%–95.8%; SGLT2i: 93.2%–95.6%). Coverage barriers remained stable in April 2022 and declined slightly in October 2022. In difference‐in‐differences analyses, the presence of a Value Statement was associated with a ~1 percentage point decline in coverage barriers for both ARNI (difference‐in‐differences estimate, −1.07% [95% CI, −1.44% to −0.70%]) and SGLT2i (−1.32% [95% CI, −1.63% to −1.00%]). Conclusions: Coverage barriers to ARNI and SGLT2i were common and changed only slightly after publication of Value Statements in HF guidelines. There is a critical need for robust strategies to improve access to life‐saving HF medications.

Objective: While timely interventions can improve medication adherence, it is challenging to identify which patients are at risk of nonadherence at point-of-care. We aim to develop and validate flexible machine learning (ML) models to predict a continuous measure of adherence to guideline-directed medication therapies (GDMTs) for heart failure (HF). Materials and Methods: We utilized a large electronic health record (EHR) cohort of 34,697 HF patients seen at NYU Langone Health with an active prescription for ≥1 GDMT between April 01, 2021 and October 31, 2022. The outcome was adherence to GDMT measured as proportion of days covered (PDC) at 6 months following a clinical encounter. Over 120 predictors included patient-, therapy-, healthcare-, and neighborhood-level factors guided by the World Health Organization’s model of barriers to adherence. We compared performance of several ML models and their ensemble (superlearner) for predicting PDC with traditional regression model (OLS) using mean absolute error (MAE) averaged across 10-fold cross-validation, % increase in MAE relative to superlearner, and predictive-difference across deciles of predicted PDC. Results: Superlearner, a flexible nonparametric prediction approach, demonstrated superior prediction performance. Superlearner and quantile random forest had the lowest MAE (mean [95% CI] = 18.9% [18.7%-19.1%] for both), followed by MAEs for quantile neural network (19.5% [19.3%-19.7%]) and kernel support vector regression (19.8% [19.6%-20.0%]). Gradient boosted trees and OLS were the 2 worst performing models with 17% and 14% higher MAEs, respectively, relative to superlearner. Superlearner demonstrated improved predictive difference. Conclusion: This development phase study suggests potential of linked EHR-pharmacy data and ML to identify HF patients who will benefit from medication adherence interventions. Discussion: Fairness evaluation and external validation are needed prior to clinical integration.