Objective:Levetiracetam is commonly prescribed for seizure prophylaxis after acute ischemic stroke (AIS) and often continued beyond discharge. While its short-term effectiveness for preventing post-stroke seizures is established, it is unclear whether prolonged use improves survival, particularly in older adults. We estimated the effect of continued levetiracetam use on 90-day mortality among Medicare beneficiaries after AIS. Methods:Using Traditional Medicare claims data (2008-2021), we identified beneficiaries aged ≥66 years hospitalized for AIS who initiated outpatient levetiracetam within 90 days of discharge. After one month of continued post-stroke use of levetiracetam (start of follow-up), we compared 90-day mortality between patients with a new levetiracetam dispensation within a 14-day grace period post-follow up and those without one. We performed cloning, censoring and weighting to address immortal time bias and estimated standardized mortality risks, risk differences, and 95% confidence intervals (CI). Results:Among 3,212 eligible beneficiaries, 1,779 (55.4%) received a new levetiracetam dispensation within the 14-day grace period. Median age was 76 years (IQR 70-83); 57.8% were female. After adjustment for demographics, hospitalization characteristics, timing of initiation, and comorbidities, continued use was associated with lower 90-day mortality than discontinuation (53 vs 62 deaths per 1,000; risk difference -9 per 1,000; 95% CI: (-12,-5)). The reduction was observed primarily among patients aged ≥75 years. Significance:Among older Medicare beneficiaries who initiated levetiracetam after AIS, continued outpatient use was associated with modestly lower 90-day mortality, particularly in those aged ≥75 years. These findings suggest potential benefits of levetiracetam continuation beyond the immediate post-stroke period.

Early detection of cognitive impairment is limited by traditional screening tools and resource constraints. We developed two large language model workflows for identifying cognitive concerns from clinical notes: (1) an expert-driven workflow with iterative prompt refinement across three LLMs (LLaMA 3.1 8B, LLaMA 3.2 3B, Med42 v2 8B), and (2) an autonomous agentic workflow coordinating five specialized agents for prompt optimization. Using Llama3.1, we optimized on a balanced refinement dataset and validated on an independent dataset reflecting real-world prevalence. The agentic workflow achieved comparable validation performance (F1 = 0.74 vs. 0.81) and superior refinement results (0.93 vs. 0.87) relative to the expert-driven workflow. Sensitivity decreased from 0.91 to 0.62 between datasets, demonstrating the impact of prevalence shift on generalizability. Expert re-adjudication revealed 44% of apparent false negatives reflected clinically appropriate reasoning. These findings demonstrate that autonomous agentic systems can approach expert-level performance while maintaining interpretability, offering scalable clinical decision supports.

ABSTRACT Efficient, accurate phenotyping for antidepressant treatment response in electronic health records (EHRs) could facilitate precision psychiatry applications but remains a challenge. Increasingly, artificial intelligence methods using “deep learning” applied to clinical data have shown promise in complex classification problems. Here, we systematically evaluate the performance of eight deep-learning-based natural language processing models in classifying response to antidepressants in a large real-world healthcare setting. We obtained data spanning 1990-2018 for adults with depression and a co-occurring antidepressant prescription from the EHR data warehouse of the Mass General Brigham healthcare system (n=111,572). Clinical notes were collected for the following time windows after antidepressant initiation: (1) 2 days to 4 weeks, (2) 4–12 weeks, and (3) 12–26 weeks. A stratified random sample of these note sets (total 4,299 across time periods) were manually reviewed to classify response status as “improved” or “no evidence of improvement” in depression symptoms. All models performed well, with areas under the receiver operator curve (AUROC) of at least 0.80. Positive predictive values (PPVs) ranged from 0.72 – 0.91. In general, models incorporating more information-dense and longer text sequences performed better than others. The best performing model (Longformer-large with sliding window) had an AUROC = 0.88 and PPV = 0.84 at a specificity of 0.88. Our results indicate that deep learning methods applied to EHR data can accurately classify antidepressant response in a real-world healthcare setting. Automated treatment response classification may facilitate a range of research and clinical decision support applications.

BackgroundApathy is a common neuropsychiatric symptom in Alzheimer's disease (AD) and has been linked to greater levels of AD biomarkers (amyloid-β, tau). However, it is unclear whether early patterns of amyloid deposition may impact development of apathy symptoms in the future.ObjectiveWe sought to examine whether amyloid-β levels, both globally and in brain regions associated with motivation, could predict future apathy symptoms.MethodsParticipants (n = 199, mean age = 79.9) were part of the Harvard Aging Brain Study, a longitudinal observational cohort of individuals without cognitive or psychiatric impairment at baseline. All underwent MRI and Pittsburgh Compound B (PiB)-PET for amyloid-β at baseline and completed questionnaires of self- and study-partner-rated apathy 7.8 ± 1 years later using the Apathy Evaluation Scale. Linear regression models assessed whether regional PiB levels predicted future apathy scores.ResultsHigher baseline cortical PiB levels in a frontal, lateral parieto-temporal, and retrosplenial aggregate were associated with greater study-partner-rated apathy, but not self-rated apathy, and no specific regional associations were observed outside of the aggregate.ConclusionsThese results provide insight into early neurobiological underpinnings of AD-related apathy. Additionally, these data may have clinical implications regarding the risk of developing apathy symptoms in amyloid-β-positive individuals as cognition declines.

Background: Early diagnosis of Alzheimer’s disease and related dementias (AD/ADRD) is critical but often constrained by limited access to fluid and imaging biomarkers, particularly in low-resource settings. Objective: To develop and evaluate a predictive model for cognitive decline using survey-based data, with attention to model interpretability and fairness. Methods: Using data from the Mexican Health and Aging Study (MHAS), a nationally representative longitudinal survey of adults aged 50 and older (N = 4095), we developed a machine learning model to predict future cognitive scores. The model was trained on survey data from 2003 to 2012, encompassing demographic, lifestyle, and social determinants of health (SDoH) variables. A stacked ensemble approach combined five base models—Random Forest, LightGBM, XGBoost, Lasso, and K-Nearest Neighbors—with a Ridge regression meta-model. Results: The model achieved a root-mean-square error (RMSE) of 39.25 (95 % CI: 38.12–40.52), representing 10.2 % of the cognitive score range, on a 20 % held-out test set. Features influencing predictions, included education level, age, reading behavior, floor material, mother’s education level, social activity frequency, the interaction between the number of living children and age, and overall engagement in activities. Fairness analyses revealed model biases in underrepresented subgroups within the dataset, such as individuals with 7–9 years of education. Discussion: These findings highlight the potential of using accessible, low-cost SDoH survey data for predicting risk of cognitive decline in aging populations. They also underscore the importance of incorporating fairness metrics into predictive modeling pipelines to ensure equitable performance across diverse groups.

OBJECTIVES:Practice effects (PEs), improvements in cognitive test performance with repeated exposure, must be addressed in longitudinal studies of cognitive aging. Although many cognitive assessments are marketed as robust to PEs, evidence is limited. Randomizing testing features enables direct quantification of PEs but remains underutilized. METHODS:Among Nurses' Health Study II participants (N=14,802), we examined PEs in the Cogstate Brief Battery arising from increased testing repetition and frequency using conditionally randomized 6- or 12-month regimens (2014 to 2019). RESULTS:Taking the assessment twice previously, compared with once previously, at the 12-month assessment (defined as frequency PEs) was associated with 0.13 SD-higher global cognitive scores (95% CI: 0.10-0.16), corresponding to between-person age differences of 4.3 to 6.8 years. Taking the second assessment 6-months compared with 12-months after baseline (defined as frequency of PEs) was associated with 0.04 SD-higher global cognitive scores ( P <0.01), corresponding to between-person 1.3 to 1.6-year age differences. Repetition and frequency PEs both appeared to be greater in magnitude for participants who were older at baseline, but uncertainty was high in formal tests of effect modification. CONCLUSIONS:Over short follow-up periods, repetition PEs may obscure age-related cognitive decline using Cogstate. Randomizing testing features can be used to strengthen cognitive aging research.

PURPOSE:Evaluating long-term effects of COVID-19 is challenging due to confounding and lack of comparable unexposed groups. The front-door criterion for causal inference identifies causal effects operating via specific pathways, even if unmeasured confounding occurs. We leverage the front-door criterion to estimate effects of COVID-19 on long-term dementia risk mediated by short-term brain magnetic resonance imaging (MRI) changes. METHODS:UK Biobank participants (n = 37,426) aged 55 + without prevalent dementia completed brain MRIs. Mediators were defined based on 31 MRI-derived variables previously shown to change due to COVID-19. We evaluated associations of these brain variables with incident all-cause dementia using confounder-adjusted Cox proportional-hazards models and combined results to infer effects of COVID-19 on dementia risk mediated via brain changes. RESULTS:Over 4.7 years (SD, 1.5) of follow-up, 112 incident dementia cases occurred. Of the 31 brain MRI-derived variables previously linked to COVID-19, 29 were associated with higher dementia incidence. Brain MRI changes of the magnitude induced on average by COVID-19 were associated with 9 % higher dementia hazard (HR: 1.09 [95 % CI, 1.06-1.12]). CONCLUSIONS:Front-door-criterion-inspired approaches can circumvent nearly intractable confounding and evaluate effects mediated by short-term physiologic changes. We find that MRI-based structural brain changes attributable to COVID-19 modestly increase dementia risk.

Background: The long preclinical phase of dementia can bias estimated effects of baseline exposures on dementia incidence. We demonstrate simulations informed by reverse Mendelian randomization (MR) findings to quantify the age-specific magnitude of reverse causation bias in analyses in observational studies of the effects of body mass index (BMI) on dementia. Methods: We simulated longitudinal trajectories of BMI and dementia risk from ages 45 to 90 years, calibrating to published evidence on age-specific dementia incidence, BMI , and associations of dementia genetic risk with BMI. Under the null that BMI does not influence dementia and an alternative that BMI at any age increases subsequent dementia risk, we simulated hypothetical cohort studies (n=20,000, average 15 years of follow-up), varying age of entry from 45 to 80 years. In each hypothetical cohort, the association of z-standardized BMI at study entry and dementia incidence were estimated using Cox proportional hazards models. Bias was quantified using the ratio of observed to true hazard ratios (RHRs). All scenarios were replicated 500 times. Results: In the absence of a causal effect of BMI on dementia, when follow-up began at age 65 years, the RHR was 0.91 (95% CI: 0.90-0.92). When follow-up began at age 80 years, the RHR decreased to 0.68 (95% CI: 0.67-0.69), indicating substantial bias attributable to reverse causation. Conclusion: Reverse causation, presumably arising from preclinical dementia, can induce substantial bias in estimates of the association between baseline exposures and dementia incidence. Simulations provide a convenient tool to quantify this bias. ### Competing Interest Statement The authors have declared no competing interest. ### Funding Statement This study was funded by NIH/NIA. ### Author Declarations I confirm all relevant ethical guidelines have been followed, and any necessary IRB and/or ethics committee approvals have been obtained. Yes I confirm that all necessary patient/participant consent has been obtained and the appropriate institutional forms have been archived, and that any patient/participant/sample identifiers included were not known to anyone (e.g., hospital staff, patients or participants themselves) outside the research group so cannot be used to identify individuals. Yes I understand that all clinical trials and any other prospective interventional studies must be registered with an ICMJE-approved registry, such as ClinicalTrials.gov. I confirm that any such study reported in the manuscript has been registered and the trial registration ID is provided (note: if posting a prospective study registered retrospectively, please provide a statement in the trial ID field explaining why the study was not registered in advance). Yes I have followed all appropriate research reporting guidelines, such as any relevant EQUATOR Network research reporting checklist(s) and other pertinent material, if applicable. Yes All data produced in the present work are contained in the manuscript

Early identification of cognitive concerns is critical but often hindered by subtle symptom presentation. This study developed and validated a fully automated, multi-agent AI workflow using LLaMA 3 8B to identify cognitive concerns in 3,338 clinical notes from Mass General Brigham. The agentic workflow, leveraging task-specific agents that dynamically collaborate to extract meaningful insights from clinical notes, was compared to an expert-driven benchmark. Both workflows achieved high classification performance, with F1-scores of 0.90 and 0.91, respectively. The agentic workflow demonstrated improved specificity (1.00) and achieved prompt refinement in fewer iterations. Although both workflows showed reduced performance on validation data, the agentic workflow maintained perfect specificity. These findings highlight the potential of fully automated multi-agent AI workflows to achieve expert-level accuracy with greater efficiency, offering a scalable and cost-effective solution for detecting cognitive concerns in clinical settings.

The digital clock drawing test (DCT) is a computerized measure examining executive functioning, information processing, and visuospatial abilities. DCT scores have been shown to discriminate well between mild cognitive impairment and Alzheimer's disease (AD) dementia and have been associated with preclinical AD pathology (i.e., amyloid, tau). Prior work has shown that amyloid moderates the relationship between depression and performance on a standard cognitive composite score. However, it is unclear whether longitudinal trajectories of depressive symptoms impact DCT scores and whether amyloid levels influence such relationships. We sought to determine whether greater depressive symptoms were associated with a decline in DCT scores and whether baseline amyloid moderated this relationship. Participants ( n  = 90, mean age=77.5±5.3; 58% female, mean education = 16.4 years) were cognitively normal (CN) and enrolled in the Harvard Aging Brain Study. The DCT (higher score=better performance) and Geriatric Depression Scale (GDS; higher score=greater depressive symptoms) were completed at baseline and annually thereafter (mean years follow-up=5.7±0.9). Ordinary least squares regression slopes were calculated for GDS data. PiB-PET (amyloid) was completed within 18 months of baseline assessments and examined using a large cortical aggregate. Separate linear mixed-effects models assessed whether ‘GDS slope*time’ or ‘GDS slope*PiB DVR*time’ predicted longitudinal DCT scores, using continuous PiB values, and controlling for age, sex, education, and random intercept/slope. GDS slope*time was not a significant predictor of longitudinal DCT scores. The 3-way-interaction with PiB was significant, such that those with greater PiB burden at baseline and higher GDS slopes showed declining DCT performance over time (beta=-8.78,95%CI[-14.66,-2.90], p  = 0.004) (Figure 1). In a cohort of CN older adults, individuals with higher baseline cortical amyloid burden and worsening depressive symptoms showed steeper declines on the DCT. These results suggest that individuals with greater/worsening depressive symptoms in the context of elevated cortical amyloid burden may be particularly vulnerable to cognitive decline (even in preclinical stages) on sensitive, easy-to-administer measures such as the DCT, similar to prior work examining cognitive testing composite scores. Findings highlight the clinical importance of monitoring emerging mood disturbances in older adults. Future work will examine how these relationships relate to clinical progression.

BACKGROUND:Compressing the duration of cognitive impairment is critical to preserve quality of life until the end. To what extent cognitive decline is compressed and cognitive resilience increases with extreme longevity is not well understood. METHODS:We used data from 13,999 deceased participants from the National Alzheimer's Coordinating Center cohort, including 8,146 with neuropathological data. Cognitive function was assessed annually (median follow-up: 4.9 years). We evaluated cognitive trajectories before death and cognitive resilience (defined as high neuropathological burden without dementia) across lifespan groups (ages 50-100+ years). RESULTS:Participants with longer lifespans, particularly centenarians, exhibited slower cognitive decline and shorter periods of cognitive impairment before death, although distinct cognitive trajectories existed among centenarians. Cognitive resilience also increased with longer lifespans, but associated factors varied. Apolipoprotein E ε2 was associated with higher cognitive resilience only in centenarians. CONCLUSION:Our findings support a general compression of cognitive decline and increased cognitive resilience in extreme longevity. HIGHLIGHTS:Individuals with longer lifespans, especially centenarians, generally exhibited better cognitive function and slower cognitive decline toward the end of life, suggesting a compression of cognitive decline in extreme longevity. Although there was a compression of cognitive decline at the group level among centenarians, heterogeneous cognitive trajectories before death were observed across individuals. The relationship between neuropathological burden and dementia risk attenuated with longer lifespans, indicating greater cognitive resilience in individuals with extreme longevity. The associations of both genetic and modifiable factors with cognitive resilience varied by lifespan.

Underdiagnosis of Alzheimer’s disease and related dementias (ADRD) leads to lost opportunities for timely intervention, increased healthcare costs, and underestimation of the true burden of disease. To address this problem, we developed an AI algorithm, Decipher-AI ( DE tection of C ognitive I mpairment PH enotypes in EHR), to screen primary care patients for undiagnosed cognitive impairment (CI). We evaluated performance across sociodemographic groups using 3 years of EHR data before the first diagnosis or most recent visit. Decipher-AI employs a two-level hierarchical model, consisting of a large language model (LLM) to generate latent representations from unstructured clinical text and a patient-level model that combines these representations with structured EHR data to predict the probability of CI (AUC: 0.98, 95% CI: 0.94, 1.00). Decipher-AI was evaluated on a test set comprising 22,000 Mass General Brigham primary care patients aged 65 years or older. The selection process involved stratified random sampling across distinct racial/ethnic strata to ensure representation of diverse sociodemographic factors (Table 1). Cognitive status labels were determined based on the presence of dementia-related diagnosis codes, while sociodemographic status was measured using the Area Deprivation Index (ADI). The AUC on the validation dataset was 0.80 [95% CI: 0.79, 0.81]; sensitivity was 0.75 [0.74, 0.77] and specificity 0.68 [0.67, 0.69], at the threshold of maximum accuracy. There were no significant differences in AUCs across sex or race/ethnic subgroups, but AUC was lower for patients under 75 years. (Figure 1A). Notably, Decipher-AI exhibited diminished performance in patients from more disadvantaged neighborhoods (Figure 1B). Further analysis revealed that these disparities were associated with fewer encounters, outpatient visits, and notes containing cognition-related keywords in patients from disadvantaged neighborhoods (Table 2). Our study highlights the potential of Decipher-AI for screening undiagnosed cognitive impairment in primary care. Nevertheless, disparities in algorithm performance underscore the importance of addressing sociodemographic inequities in EHR data. In the future, systematic tailored screening in primary care via standardized questionnaires coupled with AI-assisted chart reviews, has the potential to address the underdiagnosis of ADRD in an equitable manner.

BACKGROUND:Alzheimer's Disease and Related Dementias (ADRD) present a significant public health challenge, emphasizing the need for timely and accurate cognitive impairment (CI) diagnosis. While electronic health records (EHRs) contain valuable cognitive health data, much of this information is embedded in unstructured clinical notes. Advances in natural language processing (NLP) and large language models (LLMs) offer promising solutions, yet the application of models like GPT-4o for CI identification and staging in EHRs remains underexplored. METHOD:We developed a GPT-4o-powered framework for CI staging, integrating data querying, feature extraction, and classification. The framework was evaluated on 1002 Medicare patients from Mass General Brigham (MGB), with expert-adjudicated labels for Cognitively Unimpaired (CU), Mild Cognitive Impairment (MCI), or Dementia. To extract clinically relevant information, the framework employed GPT-4o to generate multi-note summaries, compared against keyword-based sentence extraction. GPT-4o was further employed for ordinal CI classification, producing a "summary of summaries" along with a confidence level for its final decision. Performance was benchmarked against three alternative models using USE and DementiaBERT embeddings. The framework also integrated structured answer templates, retrieval-augmented generation (RAG), and CDR domain counts with confidence levels for automated Clinical Dementia Rating (CDR) scoring, using 769 visit notes from the Massachusetts General Hospital (MGH) memory clinic. Evaluation metrics included weighted Cohen's kappa, Spearman's Rank Correlation, and Baccianella's MSE. RESULT:The framework demonstrated high accuracy in CI staging (weighted Cohen's kappa = 0.95, Spearman correlation = 0.93, Baccianella's MSE = 0.02), outperforming traditional feature extraction and embedding-based models. A confidence level stratification analysis showed that GPT-4o excelled in cases it rated with high confidence. For CDR scoring, domain counts with confidence levels yielded the best results (weighted Cohen's kappa = 0.83). CDR domain documentation in the notes significantly predicted GPT-4o's confidence in assigning global CDR. CONCLUSION:Our GPT-4o-powered framework achieves high performance in CI classification, outperforming traditional embedding models and demonstrating potential for automated chart review. However, clinical deployment requires careful consideration of misclassification risks, making a human-in-the-loop approach essential for reliability and safety. This hybrid model underscores AI's role in dementia diagnosis while ensuring interpretability and risk mitigation in real-world applications.

Clinician communication at the time of a dementia diagnosis inadequately addresses patient and caregiver needs. We aimed to characterize the disclosure experiences of patients and caregivers affected by dementia to identify communication domains that will be incorporated into the co-design of a diagnostic disclosure communication intervention. We conducted thematic analysis of individual interviews of patients and caregivers using a conceptual framework for person-centered communication. Participants included 6 patients with dementia and 15 caregivers of persons with dementia recruited from the community (n = 21; n = 17 female (81%); n = 13 Caucasian (61%); n = 4 Black or African American (19%); n = 4 Latino/a (19%); n = 2 Asian). We identified five themes. First, perceptions of respectful or disrespectful communication affects the relationship with clinicians and contributes to positive or negative experiences before, during, and after disclosure of a dementia diagnosis. Second, participants described the emotional impact of sudden or unsupported disclosures, in which they felt unprepared to receive the news or emotionally abandoned after diagnosis. Third, the absence of, or ambiguity around, a definitive dementia diagnosis contributes to distress and feeling dismissed by clinicians. Fourth, mixed responses to recommendations after disclosure reveals the need for a personalized care planning process. Fifth, considerations around the timing of prognostic communication and advance care planning are necessary to meet the needs of individuals with different emotional readiness, information preferences, and cultural beliefs. Dementia diagnostic disclosure would benefit from a tailored, structured, and incremental approach that prioritizes respectful communication, emotional support, and personalized care planning to meet the needs of patients and caregivers.

Purpose: Sexual minority (SM) women have more dementia risk factors than heterosexual women, but it remains unknown whether they experience increased symptoms of subjective cognitive decline (SCD)-a key predictor of dementia. Methods: We investigated sexual orientation-related disparities in SCD in Nurses' Health Study II (N = 70,772). Sexual orientation subgroups included completely heterosexual (n = 62,884); participants identifying as heterosexual with same-sex experience ("heterosexual-SM", n = 5017); and participants identifying as mostly heterosexual (n = 1825), bisexual (n = 287), or lesbian/gay (n = 759). SCD was measured using seven symptoms from the Structured Telephone Interview for Dementia Assessment, controlling for demographics with Poisson regression models. Results: Relative to completely heterosexual participants, SM participants had 29% more SCD symptoms (95% confidence interval [CI] = 1.26-1.32). Symptoms were elevated in every SM subgroup; the largest disparities were among bisexual and mostly heterosexual subgroups (adjusted risk ratios for 1-unit increment in symptoms [aRR]: 1.60, 95% CI = 1.45-1.77; 1.48, 95% CI = 1.42-1.54, respectively) followed by lesbian/gay (aRR: 1.22, 95% CI = 1.14-1.31) and heterosexual-SM participants (aRR: 1.21, 95% CI = 1.18-1.25). Conclusion: SM women-particularly bisexual and mostly heterosexual women-had more symptoms of SCD than completely heterosexual women. These findings align with known sexual orientation-related disparities in dementia risk factors (e.g., mental health, substance use), and indicate that better understanding and closer monitoring of cognitive health in SM groups remains important for prevention efforts as an increasing proportion of aging Americans identifies as SM.

BACKGROUND:Early diagnosis of Alzheimer's disease and related dementias (AD/ADRD) is essential for timely intervention, but many current screening methods, such as blood-based biomarkers and imaging techniques, remain inaccessible. Social determinants of health (SDoH) and lifestyle factors offer readily available alternative approach to predicting cognitive decline. Thus, we developed a stacking ensemble model that uses SDoH data for early detection of cognitive impairment. METHOD:Our model was trained using data from the Mexican Health and Aging Study (MHAS), a national longitudinal survey of adults aged 50 and older in Mexico (n = 4,095). The dataset includes variables covering demographics, economic conditions, self-reported health, and lifestyle behaviors collected in 2003 and 2012. The model employs a two-level stacking ensemble architecture to predict global cognition in 2016 and 2021 (Figure 1) (1) First-layer models provide feature representations; (2) Second-layer model integrates predictions from the first-layer models to predict the cognitive composite score. RESULT:The ensemble model achieved a Root-Mean-Square Deviation (RMSE) of 39.25 on the test set (Figure 2a), corresponding to an average deviation of approximately 10.2% from the full range of the cognitive score (0 to 384). Among the base models, LightGBM demonstrated the best performance (Figure 2b). Feature importance analysis highlighted the key social and lifestyle factors influencing cognitive function (Figure 2c). Sensitivity analyses showed that perturbations to top-ranked features caused minimal RMSE fluctuations, demonstrating model robustness (Figure 2d). Additionally, model performance analysis revealed biases in certain education and social engagement subgroups, potentially because of lower representation of those subgroups in the dataset (Figure 3). CONCLUSION:Our study demonstrates the feasibility of using readily available SDOH data for early detection of cognitive impairment. The model provides a balance between predictive performance and explainability. Future work should predict change in cognitive scores since cognitive scores are influenced by social determinants of health.

This cohort study describes the monthly and annual population-based cardiac mortality rates during and after the COVID-19 pandemic.

Fall-related injuries (FRIs) are a major cause of hospitalizations among older patients, but identifying them in unstructured clinical notes poses challenges for large-scale research. In this study, we developed and evaluated natural language processing (NLP) models to address this issue. We utilized all available clinical notes from the Mass General Brigham health-care system for 2100 older adults, identifying 154 949 paragraphs of interest through automatic scanning for FRI-related keywords. Two clinical experts directly labeled 5000 paragraphs to generate benchmark-standard labels, while 3689 validated patterns were annotated, indirectly labeling 93 157 paragraphs as validated-standard labels. Five NLP models, including vanilla bidirectional encoder representations from transformers (BERT), the robustly optimized BERT approach (RoBERTa), ClinicalBERT, DistilBERT, and support vector machine (SVM), were trained using 2000 benchmark paragraphs and all validated paragraphs. BERT-based models were trained in 3 stages: masked language modeling, general boolean question-answering, and question-answering for FRIs. For validation, 500 benchmark paragraphs were used, and the remaining 2500 were used for testing. Performance metrics (precision, recall, F1 scores, area under the receiver operating characteristic curve [AUROC], and area under the precision-recall [AUPR] curve) were employed by comparison, with RoBERTa showing the best performance. Precision was 0.90 (95% CI, 0.88-0.91), recall was 0.91 (95% CI, 0.90-0.93), the F1 score was 0.91 (95% CI, 0.89-0.92), and the AUROC and AUPR curves were [both??] 0.96 (95% CI, 0.95-0.97). These NLP models accurately identify FRIs from unstructured clinical notes, potentially enhancing clinical-notes-based research efficiency.

Timely identification of individuals at-risk for Alzheimer's disease (AD) is pivotal for secondary prevention and requires innovative approaches. Combining remote smartphone-based cognitive assessments with blood-based biomarkers holds promise for both sensitive and scalable detection of early AD-related cognitive changes. Here, we aimed to investigate whether the Boston Remote Cognitive Assessment of NeuroCognitive Health (BRANCH) captures cognitive changes associated with early AD pathophysiology as measured by plasma p -tau217. N  = 254 cognitively unimpaired older adults (age=74.5±8.6, 68% female, 16.6±2.4 years of education) from four well-characterized cohorts completed multi-day BRANCH on their personal device. Multiday BRANCH includes two associative memory tests (Face Name and Groceries Prices) and a processing speed test with an associative memory component (Digit Signs) with identical stimuli repeated for seven consecutive days. For each test, an MDLC score was computed using an area under the curve method combining day 1 performance with a non-linear learning trajectory over the subsequent six days. MDLCs for each individual test were averaged into a BRANCH Composite MDLC. All cohorts had standardized in-clinic cognitive test data available, from which a Preclinical Alzheimer's Cognitive Composite (PACC-5) score was derived. Concentrations of plasma p -tau217 were measured using the Meso Scale Discovery platform. Linear regression models adjusting for age, sex, years of education and study cohort were used to investigate the association between p -tau217 (log-transformed values) and BRANCH MDLC scores. For comparison, similar analyses were run with PACC-5 scores and p -tau217. Lower BRANCH Composite MDLC scores were associated with higher p -tau217 levels (corrected std. β = -0.23, 95%CI [-0.44 – -0.02], p  = 0.031) (Figure 1), which was primarily driven by the Digit Signs test (corrected std. β = -0.22, 95%CI [-0.42 – -0.02], p  = 0.031). In contrast, we did not find an association between the PACC-5 and p -tau217 (corrected std. β = -0.15, 95%CI [-0.37 – 0.07], p  = 0.187). These results complement our previous work that multi-day BRANCH may improve the detection of very subtle memory deficits that are associated with early AD pathophysiology. Combining a remote and sensitive cognitive paradigm like BRANCH with plasma biomarkers may facilitate scalable detection of those at risk for AD-related cognitive decline.