About

Erjia Cui is an Assistant Professor in the Division of Biostatistics and Health Data Science at the University of Minnesota. She earned her PhD in Biostatistics from Johns Hopkins University in 2023 under the supervision of Ciprian Crainiceanu. Prior to her doctoral studies, she obtained a Bachelor's degree in Statistics from Zhejiang University in 2018. Her research is inspired by the complex data structures arising from large cohort studies such as the UK Biobank and NHANES. She focuses on developing Functional Data Analysis methods along with scalable and reproducible software implementations. A major direction of her scientific research is to quantify the role of objectively measured physical activity collected from wearable devices on aging, disease progression, and mortality.

Research topics

• Computer Science
• Artificial Intelligence
• Machine Learning
• Mathematics
• Data Mining
• Econometrics
• Biology
• Computational biology
• Statistics
• Medicine

Selected publications

• Capturing instantaneous neural signal-behavior relationships with concurrent functional mixed models

  eLife · 2026-03-18

  articleOpen access

  We previously proposed an analysis framework for fiber photometry data based on functional linear mixed models (FLMMs). Functional LMMs allow modeling associations between photometry traces and trial-specific scalar values like behavioral summaries and session number, while also accounting for between-animal heterogeneity. Here, we extend the method to concurrent FLMMs (cFLMMs), a method that can fit the instantaneous relationship between functional outcomes and functional covariates. Concurrent FLMMs enable testing of how the photometry signal is associated with, for example, a behavioral variable that evolves across within-trial timepoints (e.g. animal speed). cFLMMs can also model the relationship between the photometry signal and covariates in experiments with variable trial lengths (e.g., in studies where trials end when an animal responds). We illustrate the application of cFLMMs on two published studies and show the method can identify signal–behavior associations in analyses not possible with FLMMs. We find that analyzing photometry–behavior associations based on behavioral summaries (e.g., latency-to-response, average lick rate) can lead to misleading conclusions. We published our method in the <monospace>fastFMM</monospace> package, available as an R package through GitHub (https://github.com/awqx/fastFMM).

  PublisherOA PDFDOI

• Association of chronic disease risk and physical activity measured by wearable devices in the All of Us program

  Communications Medicine · 2026-01-14 · 1 citations

  articleOpen access

  Physical activity plays an important role in preventing chronic diseases, but most studies rely on self-reported or short-term data that fail to capture habitual behavior. This study utilizes Fitbit data to investigate the relationship between physical activity and various chronic diseases. We analyzed data from 22,019 participants in the All of Us Research Program who shared at least six months of Fitbit activity data linked with electronic health records. Various physical activity patterns were evaluated using Cox proportional hazards and logistic regression models, adjusting for age, sex, and body mass index (BMI). To test robustness, sensitivity analyses were conducted using obesity defined by BMI, applying a two-year exclusion window for outcome diagnoses to mitigate potential reverse causation, and incorporating lifestyle covariates (smoking and alcohol use) under a simplified directed acyclic graph (DAG) framework to address residual confounding. Here, we show that higher physical activity levels are associated with lower risks of multiple chronic conditions. Higher daily step counts were negatively associated with obesity and type 2 diabetes, while greater elevation gains and longer vigorous activity are associated with lower risks of conditions such as morbid obesity, obstructive sleep apnea, and major depressive disorder. All sensitivity analyses yield consistent results, supporting the robustness of findings against reverse causation and lifestyle confounding. Higher physical activity and lower sedentary time may help prevent diverse chronic diseases. These findings demonstrate the potential of large-scale wearable data to inform personalized prevention and population health strategies. This study looked at how physical activity affects the risk of developing long-term health problems. Researchers used data from people who wore Fitbit devices to track their daily activity, including step count, exercise intensity, and how long they were active or sitting. These data were linked with each person’s medical records to understand how activity levels related to chronic diseases such as obesity, diabetes, and depression. The study found that people who were more active—taking more steps, climbing more elevation, and spending more time in intense activity—had a lower chance of developing these diseases. The results suggest that being more active and sitting less can help people stay healthier and may support future public health advice and personal health planning. Hou et al analyze Fitbit and electronic health record data from over 22,000 participants in the All of Us Research Program to examine links between daily physical activity and chronic disease risk. Higher activity and lower sedentary time are associated with reduced risks of obesity, diabetes, and other chronic conditions.

  PublisherOA PDFDOI

• Capturing instantaneous neural signal-behavior relationships with concurrent functional mixed models

  eLife · 2026-03-18

  articleOpen access

  We previously proposed an analysis framework for fiber photometry data based on functional linear mixed models (FLMMs). Functional LMMs allow modeling associations between photometry traces and trial-specific scalar values like behavioral summaries and session number, while also accounting for between-animal heterogeneity. Here, we extend the method to concurrent FLMMs (cFLMMs), a method that can fit the instantaneous relationship between functional outcomes and functional covariates. Concurrent FLMMs enable testing of how the photometry signal is associated with, for example, a behavioral variable that evolves across within-trial timepoints (e.g. animal speed). cFLMMs can also model the relationship between the photometry signal and covariates in experiments with variable trial lengths (e.g., in studies where trials end when an animal responds). We illustrate the application of cFLMMs on two published studies and show the method can identify signal–behavior associations in analyses not possible with FLMMs. We find that analyzing photometry–behavior associations based on behavioral summaries (e.g., latency-to-response, average lick rate) can lead to misleading conclusions. We published our method in the <monospace>fastFMM</monospace> package, available as an R package through GitHub (https://github.com/awqx/fastFMM).

  PublisherOA PDFDOI

• Abstract 67: Objectively Measured vs. Self-Reported Physical Activity and Coronary Artery Calcification: The Atherosclerosis Risk in Communities Study

  Circulation · 2026-03-24

  article

  Introduction: Coronary artery calcification (CAC) is a marker of subclinical atherosclerosis with important implications for cardiovascular risk assessment. Despite established cardiovascular benefits of physical activity, its relationship with CAC remains unclear. Hypothesis: The CAC Agatston score has an inverse association with both objectively measured and self-reported physical activity. Methods: We analyzed data from 1,449 participants in the Atherosclerosis Risk in Communities (ARIC) study to investigate the cross-sectional association of objective and self-reported physical activity at visit 6 (2016-17) with Agatston scores from cardiac CT at visit 7 (2018-19). Objective physical activity was measured using an accelerometer embedded in the Zio® XT ECG monitor, which provided continuous monitoring for up to 14 days. Self-reported activity in the past year was assessed via modified Baecke Questionnaire. Multivariable linear regression and logistic regression were used for analysis. Results: Objective and self-reported physical activity assessments were obtained at a median (IQR) age of 78 (75-81) years; 60.4% were females, and 23.1% were Black. The Spearman correlation coefficient between objectively measured and self-reported average daily hours of moderate-to-vigorous physical activity (MVPA) was 0.40. Cardiac CT scan was performed at a median (IQR) of 1.7 (1.5-2.0) years thereafter. Overall, the Agatston score was 0 in 10.1% and was > 400 in 39.1% of the participants. We observed a significant J-shaped association between objectively measured MVPA and Agatston score ( Figure 1 ). The estimated mean Agatston score was highest at 64 (at 0 hours/day of MVPA) and reached its nadir of 30 at 0.74 hours/day of MVPA. Beyond 0.74 hours/day of MVPA, the estimated mean Agatston score showed a slight upward trend. Similarly, among participants with detectable CAC (Agatston score > 0), 0.80 hours/day of MVPA corresponded to the lowest marginal predicted probability (17%) of severe CAC (Agatston score > 400) ( Figure 2 ). In contrast, there was no significant association between self-reported MVPA and Agatston scores. Conclusions: Objectively measured physical activity demonstrated a J-shaped relationship with CAC, with the lowest Agatston score observed at approximately 45 minutes daily of MVPA. Self-reported activity may inadequately capture this association, highlighting the importance of objective assessment in cardiovascular risk evaluation.

  PublisherDOI

• Estimating causal effects of functional treatments with modified functional treatment policies

  ArXiv.org · 2026-02-09

  articleOpen access

  Functional data are increasingly prevalent in biomedical research. While functional data analysis has been established for decades, causal inference with functional treatments remains largely unexplored. Existing methods typically focus on estimating the causal average dose response functional (ADRF), which requires strong positivity assumptions and offers limited interpretability. In this work, we target a new causal estimand, the modified functional treatment policy (MFTP), which focuses on estimating the average potential outcome when each individual slightly modifies their treatment trajectory from the observed one. A major challenge for this new estimand is the need to define an average over an infinite-dimensional object with no density. By proposing a novel definition of the population average over a functional variable using a functional principal component analysis (FPCA) decomposition, we establish the causal identifiability of the MFTP estimand. We further derive outcome regression, inverse probability weighting, and doubly robust estimators for the MFTP, and provide theoretical guarantees under mild regularity conditions. The proposed estimators are validated through extensive simulation studies. Applying our MFTP framework to the National Health and Nutrition Examination Survey (NHANES) accelerometer data, we estimate the causal effects of reducing disruptive nighttime activity and low-activity duration on all-cause mortality.

  PublisherOA PDF

• Abstract TP055: ICAS-Related Stroke: Remote Tele-Based Care Matches or Exceeds In-Person Best Practice Adherence

  Stroke · 2026-01-29

  article

  Background: Remote tele-based stroke evaluation and treatment programs are a promising option for underserved populations. Implementation and adherence to guideline-based best practices among these programs remains relatively unexplored. We aimed to characterize adherence to best practices in intracranial atherosclerotic stenosis (ICAS)-related stroke management between in-person versus post-acute telestroke, both managed by comprehensive stroke center (CSC) providers. Methods: We performed a retrospective cohort review, identifying patients with ICAS-related stroke across the MHealth Fairview system. The system comprises 9 hospitals, including 2 CSCs with in-person stroke team coverage and 7 hospitals with remote tele-based coverage. Adherence to best practices was determined using four primary outcome measures including rates of permissive hypertension within the first 48 hours of hospitalization, high-intensity statin prescription, time to initiation of antiplatelet medication(s), and appropriate antithrombotic therapy following stroke. Statistical tests included Pearson’s Chi-squared and Fisher’s exact tests (categorical outcomes) or Wilcoxin rank sum test (continuous outcomes). Results: Among 112 patients, 79 were evaluated and treated by in-person CSC providers. The remaining 33 were evaluated and treated remotely by the same CSC provider group. Demographic characteristics and neuroimaging findings were similar between inpatient versus tele-based study participants (Table 1). Adherence to best practices was similar between in-person versus tele-based coverage including rates of permissive hypertension (63.3% vs 56.3%, p=0.52) and appropriate antithrombotic treatment (96% vs 97%, p=1.00). When compared to patients at in-person sites, patients at telestroke sites were significantly more likely to receive high intensity statin (84.8% vs 51.9%, p=0.002). Patients at in-person sites had a significantly longer time from presentation to time of first aspirin dose (median 6.75 hours [IQR 3-19] vs 4.0 hours [2.25-7.25], p=0.047) and to first clopidogrel dose (10 hours [IQR 4-27] vs 4.5 hours [IQR 2.25-12.75], p=0.02) when compared to tele-based sites. Conclusions: Guideline appropriate care was provided at similar or better rates at sites with tele-based rounding versus in-person rounding. While promising, continued evaluation of adherence to best practices within tele-based care settings remains crucial to ensure sustainable success of similar programs.

  PublisherDOI

• Estimating causal effects of functional treatments with modified functional treatment policies

  Open MIND · 2026-02-09

  preprint

  Functional data are increasingly prevalent in biomedical research. While functional data analysis has been established for decades, causal inference with functional treatments remains largely unexplored. Existing methods typically focus on estimating the causal average dose response functional (ADRF), which requires strong positivity assumptions and offers limited interpretability. In this work, we target a new causal estimand, the modified functional treatment policy (MFTP), which focuses on estimating the average potential outcome when each individual slightly modifies their treatment trajectory from the observed one. A major challenge for this new estimand is the need to define an average over an infinite-dimensional object with no density. By proposing a novel definition of the population average over a functional variable using a functional principal component analysis (FPCA) decomposition, we establish the causal identifiability of the MFTP estimand. We further derive outcome regression, inverse probability weighting, and doubly robust estimators for the MFTP, and provide theoretical guarantees under mild regularity conditions. The proposed estimators are validated through extensive simulation studies. Applying our MFTP framework to the National Health and Nutrition Examination Survey (NHANES) accelerometer data, we estimate the causal effects of reducing disruptive nighttime activity and low-activity duration on all-cause mortality.

  DOI

• External validation of the predictive swallow score for dysphagia in stroke patients

  Journal of Stroke and Cerebrovascular Diseases · 2026-02-12

  articleOpen access

  INTRODUCTION: Post-stroke dysphagia (PSD) is a common complication following acute ischemic stroke (AIS). Predicting the recovery of swallow function remains challenging. The Predictive Swallow Score (PRESS) model, derived and validated in a Swiss cohort, sought to predict the recovery of PSD after AIS. We aimed to validate the PRESS model in a US-cohort, conducting a two-center retrospective review of 149 patients with AIS and functional oral intake scale (FOIS) ≤ 4. METHODS: We collected the predictors of recovery of PSD according to PRESS (age, NIH Stroke Scale (NIHSS), any2 score, stroke location, FOIS score), with a primary outcome of impaired swallow at day 7 (FOIS ≤ 4). Model validation was completed using the Hosmer-Lemeshow (HL) test, calibration plots, and AUC analysis. RESULTS: =48.343, df=5), and the calibration curve analysis (intercept = -0.80 (95% CI: -1.21 to -0.38), slope = 0.60 (95% CI: 0.37 to 0.82)) also demonstrated a poor calibration of the model. Area under the curve analyses demonstrated a C statistic of 0.75 (95% CI 0.67-0.82), indicating suboptimal model discrimination in predicting the recovery of swallow 7 days following AIS. In particular, the model overpredicted dysphagia severity at day-7 in patients with higher PRESS scores and more severe strokes. CONCLUSION: Further validation of the PRESS score in prospective cohorts is warranted. The suboptimal model performance could be attributed to temporal advances in stroke care, as the original PRESS cohort was derived between 2011 and 2014. Geographic variability in acute stroke care practice could also be a factor, as the PRESS score was derived solely from a European cohort. This study, however, is limited by its retrospective design and a lack of generalizability.

  PublisherDOI

• Capturing instantaneous neural signal-behavior relationships with concurrent functional mixed models

  bioRxiv (Cold Spring Harbor Laboratory) · 2025-09-12

  preprintOpen access

  Abstract We previously proposed an analysis framework for fiber photometry data based on functional linear mixed models (FLMMs). Functional LMMs allow modeling associations between photometry traces and trial-specific scalar values like behavioral summaries and session number, while also accounting for between-animal heterogeneity. Here, we extend the method to concurrent FLMMs (cFLMMs), a method that can fit the instantaneous relationship between functional outcomes and functional covariates. Concurrent FLMMs enable testing of how the photometry signal is associated with, for example, a behavioral variable that evolves across within-trial timepoints (e.g. animal speed). cFLMMs can also model the relationship between the photometry signal and covariates in experiments with variable trial lengths (e.g., in studies where trials end when an animal responds). We illustrate the application of cFLMMs on two published studies and show the method can identify signal–behavior associations in analyses not possible with FLMMs. We find that analyzing photometry–behavior associations based on behavioral summaries (e.g., latency-to-response, average lick rate) can lead to misleading conclusions. We published our method in the fastFMM package, available as an R package through GitHub ( https://github.com/awqx/fastFMM ).

  PublisherOA PDFDOI

• Suicidal Ideation is Associated with Greater Autonomic and Motor Dysfunction in Patients with REM Sleep Behavior Disorder (S6.009)

  Neurology · 2025-04-07

  article

  We aim to assess markers of neurodegenerative disease burden in patients with REM sleep behavior disorder (RBD) who endorsed suicidal ideation in the North American Prodromal Synucleinopathy (NAPS) consortium registry.

  PublisherDOI

Frequent coauthors

• Ciprian M. Crainiceanu

  Johns Hopkins University

  54 shared

• Andrew Leroux

  University of Colorado Anschutz Medical Campus

  53 shared

• Martin A. Lindquist

  28 shared

• Qier Meng

  Johns Hopkins University

  25 shared

• Ellen M. Mowry

  Johns Hopkins Medicine

  25 shared

• Paul W. Blair

  Henry M. Jackson Foundation

  12 shared

• Jeff Goldsmith

  11 shared

• Gabriel Loewinger

  National Institute on Alcohol Abuse and Alcoholism

  11 shared

Education

• PhD Candidate, Biostatistics

  Johns Hopkins University

• Bachelor of Science, Statistics

  Zhejiang University

  2018

Awards & honors

• Member, Alpha Chapter, Delta Omega Honorary Society in Publi…