Elizabeth A. Chrischilles

Background: Acute kidney injury (AKI) is a common, life-threatening complication among hospitalized patients. Timely prediction is crucial for prevention, while accurate prognosis following onset is essential for guiding care. Although electronic health records (EHRs) provide rich longitudinal data, their high dimensionality, heterogeneity, and temporal complexity pose challenges for traditional machine learning. Objective: The aim of this study was to develop and evaluate a transformer-based model for generalizable AKI prediction and prognosis by modeling long-range temporal dependencies and complex relationships in heterogeneous EHR data and to examine its interpretability. Methods: We developed the Biomarker-Aware Transformer for AKI (BAT-AKI), which integrates multidimensional embeddings capturing temporal dynamics, textual descriptions, hierarchical ontology structures, and semantic relationships in EHRs. To enhance sensitivity to AKI-related biomarker fluctuations, BAT-AKI combines masked language modeling with a novel biomarker abnormality-aware pretraining task. The model was pretrained and fine-tuned for 4 tasks: AKI onset, early reversal, recovery at discharge, and 90-day post-AKI mortality. Pretraining and internal validation used data from the National Patient-Centered Clinical Research Network (PCORnet) site, with external validation across 3 additional sites. Results: BAT-AKI achieved the strongest overall discrimination, outperforming baseline models in area under the receiver operating characteristic curve (AUROC) and area under the precision-recall curve (AUPRC) on most tasks. For AKI onset prediction, it achieved an AUROC of 0.8549 (95% CI 0.8433-0.8658) and an AUPRC of 0.6041 (95% CI 0.5786-0.6292), exceeding the second-best baseline by 2.8% in AUROC and 8.1% in AUPRC. Across prognostic tasks, BAT-AKI achieved AUROCs of 0.6006 (95% CI 0.5703-0.6318), 0.6080 (95% CI 0.5745-0.6397), and 0.8017 (95% CI 0.7708-0.8311) and AUPRCs of 0.7036 (95% CI 0.6690-0.7388), 0.7417 (95% CI 0.7082-0.7765), and 0.5063 (95% CI 0.4422-0.5684) for early reversal, recovery at discharge, and 90-day post-AKI mortality, respectively, corresponding to relative gains of 2.4% to 6% in AUROC and 3.7% to 4.1% in AUPRC over the second-best baseline. Performance remained stable across external cohorts, with AKI onset AUROCs ranging from 0.7754 (95% CI 0.7677-0.7827) to 0.7989 (95% CI 0.7885-0.8092) and AUPRCs ranging from 0.4777 (95% CI 0.4631-0.4924) to 0.5687 (95% CI 0.5495-0.5871), corresponding to relative improvements of 3.7% to 5.3% in AUROC and 16.8% to 25.3% in AUPRC over the transformer baseline. In a low-resource setting using only 2% of the fine-tuning data, BAT-AKI achieved an AUROC of 0.7898 (95% CI 0.7759-0.8036) and an AUPRC of 0.4903 (95% CI 0.4640-0.5178) for AKI onset prediction, exceeding the second-best baseline by 4.2% in AUROC and 10.9% in AUPRC. Visualization of embeddings, attention patterns, and integrated-gradient feature attributions highlighted clinically meaningful relationships and patient-specific trajectories. Conclusions: BAT-AKI provides generalizable AKI prediction and prognosis with consistent discrimination across institutions, while offering interpretability insights into its predictions, supporting its potential for early warning and AKI care optimization.

BACKGROUND: Oral anticoagulation (OAC) remains underused in patients with atrial fibrillation or flutter at high risk of ischemic stroke. We aimed to characterize documented OAC prescribing at diagnosis and during 1-year follow-up and identify factors associated with baseline OAC prescribing. METHODS: Using electronic health records from 7 US health systems participating in the National Patient-Centered Clinical Research Network between 2015 and 2019, we identified patients with newly diagnosed atrial fibrillation at high risk of ischemic stroke (CHA ₂ DS ₂ -VASc score ≥2 for men or ≥3 for women). The primary outcomes were documented OAC prescription at baseline and during 1-year follow-up. Exposures included demographic, clinical, and health-system characteristics. Multivariable logistic regression was used to identify factors associated with baseline OAC prescribing and loss of documented OAC prescribing during follow-up. RESULTS: Among 161 295 patients (median age, 76 years [interquartile range, 68–82]; 70 327 [43.6%] women), 92 518 (57.4%) had a documented OAC prescription at baseline, increasing from 56.6% in 2015 to 62.6% in 2019. Among those with a documented baseline OAC prescription, 45.6% had no documented OAC prescription during 1-year follow-up. Among 68 777 patients without a documented baseline OAC prescription, 12.5% had an OAC prescription documented during follow-up. Dementia (odds ratio, 0.58 [95% CI, 0.56–0.61]), prior major bleeding (odds ratio, 0.69 [95% CI, 0.63–0.75]), and renal disease (odds ratio, 0.84 [95% CI, 0.81–0.88]) were associated with lower odds of baseline OAC prescribing, whereas heart failure was associated with higher odds (odds ratio, 1.40 [95% CI, 1.37–1.43]). CONCLUSIONS: Among patients with newly diagnosed atrial fibrillation at high risk for ischemic stroke, OAC prescribing at diagnosis was suboptimal, and loss of documented OAC prescribing was common during the first year. Baseline OAC use was more strongly associated with bleeding risk, frailty, and comorbidity burden, and less consistently aligned with ischemic stroke risk.

Background: Reductions in blood pressure (BP) control among patients with hypertension were observed early in the COVID‐19 pandemic. The degree to which BP control may have returned to prepandemic levels is unknown. Methods: Individuals aged 18 to 85 years with hypertension from 17 health systems participating in the National Patient‐Centered Clinical Research Network were identified using electronic health record data collected as part of routine care. BP control (percentage of patients whose most recent BP measurement was <140/<90 mm Hg) was estimated in a series of 12‐month rolling measurement periods from 2017 through 2022 (January 1, 2017 through December 31, 2017; April 1, 2017 through March 31, 2018; … January 1, 2022 through December 31, 2022). Differences in average BP control between 2022 (January 1, 2022 through December 31, 2022) and 2019 (January 1, 2019 through December 31, 2019) were estimated overall (adjusted for age, sex, and race and ethnicity) and by race and ethnicity (adjusted for age and sex). Results: Our sample included 1 193 314 persons with hypertension in 2019 (48.9% aged 65–85 years, 52.9% men, 66.2% non‐Hispanic White) and 1 499 418 individuals in 2022 (50.6% aged 65–85 years, 47.1% men, 62.7% non‐Hispanic White). The weighted average BP control dropped from 65.3% in 2019 to 61.8% in 2020 and then partially recovered to 62.6% in 2022 (adjusted mean difference, −2.6 percentage points [95% CI, −5.0 to −0.2]). Non‐Hispanic Asian individuals experienced the largest temporal drop in BP control, declining from 68.4% in 2019 to 63.9% in 2022. Conclusions: BP control was disrupted during the COVID‐19 pandemic and had not fully rebounded to prepandemic levels by the end of 2022. Continued surveillance is needed to determine whether the decline in BP control will persist and will result in future adverse cardiovascular events.

Background: Neuroendocrine tumors (NETs) are rare, heterogeneous neoplasms associated with prolonged survival and substantial symptom burden. However, patient-reported outcomes across NET subtypes remain poorly characterized, particularly in real-world settings. This study describes baseline health-related quality of life (HRQoL) and care experiences among patients with gastroenteropancreatic and lung NETs, examining differences by tumor site and time since diagnosis. Methods: The Neuroendocrine Tumors–Patient Reported Outcomes study is a prospective, multi-institutional US cohort of adults (aged 18 years and older) with incident small intestinal, pancreatic, gastroenteropancreatic, or lung NETs diagnosed from January 2018 through September 2024, identified via a validated electronic medical record–based computable phenotype. Baseline surveys assessed HRQoL, symptoms, care experiences, and clinical characteristics using validated instruments. Descriptive statistics and standardized mean differences compared responses by NET site and time since diagnosis. Results: Among 2367 participants (mean age = 57.8 years; 57.3% female), 1974 had gastroenteropancreatic NETs (659 small intestinal NET, 555 pancreatic NET) and 393 had lung NETs. Fatigue (mean = 33.0), insomnia (mean = 32.5), and diarrhea (mean = 25.7) were the most burdensome symptoms. Lung NET patients reported worse dyspnea (standardized mean difference = 0.58, P < .001) and lower physical, role, and global QoL scores than those with gastroenteropancreatic NETs, while pancreatic NET patients reported better functioning. Diarrhea worsened over time, especially in small intestinal NETs. Most rated care highly (75.3%) but cited concerns about treatment side effects (80.4%), costs (60.7%), and travel burden (58.8%). Conclusions: This large US cohort reveals persistent symptom burden and HRQoL variation by tumor site and disease duration, underscoring the need for longitudinal HRQoL assessment in NET care.

BACKGROUND Acute kidney injury (AKI) is a common, life-threatening complication among hospitalized patients. Timely prediction is crucial for prevention, while accurate prognosis following onset is essential for guiding care. Although electronic health records (EHRs) provide rich longitudinal data, their high dimensionality, heterogeneity, and temporal complexity pose challenges for traditional machine learning. OBJECTIVE The aim of this study was to develop and evaluate a transformer-based model for generalizable AKI prediction and prognosis by modeling long-range temporal dependencies and complex relationships in heterogeneous EHR data and to examine its interpretability. METHODS We developed the Biomarker-Aware Transformer for AKI (BAT-AKI), which integrates multidimensional embeddings capturing temporal dynamics, textual descriptions, hierarchical ontology structures, and semantic relationships in EHRs. To enhance sensitivity to AKI-related biomarker fluctuations, BAT-AKI combines masked language modeling with a novel biomarker abnormality-aware pretraining task. The model was pretrained and fine-tuned for 4 tasks: AKI onset, early reversal, recovery at discharge, and 90-day post-AKI mortality. Pretraining and internal validation used data from the National Patient-Centered Clinical Research Network (PCORnet) site, with external validation across 3 additional sites. RESULTS BAT-AKI achieved the strongest overall discrimination, outperforming baseline models in area under the receiver operating characteristic curve (AUROC) and area under the precision-recall curve (AUPRC) on most tasks. For AKI onset prediction, it achieved an AUROC of 0.8549 (95% CI 0.8433-0.8658) and an AUPRC of 0.6041 (95% CI 0.5786-0.6292), exceeding the second-best baseline by 2.8% in AUROC and 8.1% in AUPRC. Across prognostic tasks, BAT-AKI achieved AUROCs of 0.6006 (95% CI 0.5703-0.6318), 0.6080 (95% CI 0.5745-0.6397), and 0.8017 (95% CI 0.7708-0.8311) and AUPRCs of 0.7036 (95% CI 0.6690-0.7388), 0.7417 (95% CI 0.7082-0.7765), and 0.5063 (95% CI 0.4422-0.5684) for early reversal, recovery at discharge, and 90-day post-AKI mortality, respectively, corresponding to relative gains of 2.4% to 6% in AUROC and 3.7% to 4.1% in AUPRC over the second-best baseline. Performance remained stable across external cohorts, with AKI onset AUROCs ranging from 0.7754 (95% CI 0.7677-0.7827) to 0.7989 (95% CI 0.7885-0.8092) and AUPRCs ranging from 0.4777 (95% CI 0.4631-0.4924) to 0.5687 (95% CI 0.5495-0.5871), corresponding to relative improvements of 3.7% to 5.3% in AUROC and 16.8% to 25.3% in AUPRC over the transformer baseline. In a low-resource setting using only 2% of the fine-tuning data, BAT-AKI achieved an AUROC of 0.7898 (95% CI 0.7759-0.8036) and an AUPRC of 0.4903 (95% CI 0.4640-0.5178) for AKI onset prediction, exceeding the second-best baseline by 4.2% in AUROC and 10.9% in AUPRC. Visualization of embeddings, attention patterns, and integrated-gradient feature attributions highlighted clinically meaningful relationships and patient-specific trajectories. CONCLUSIONS BAT-AKI provides generalizable AKI prediction and prognosis with consistent discrimination across institutions, while offering interpretability insights into its predictions, supporting its potential for early warning and AKI care optimization. CLINICALTRIAL

Misclassification and measurement error is a significant issue in epidemiologic research. When utilizing large datasets like the UK Biobank in particular, we found that pancreatic neuroendocrine neoplasms are often erroneously included in studies focused on exocrine pancreatic cancer. This happens when studies use only International Classification of Diseases codes to define their cases. The use of International Classification of Diseases for Oncology, Third Edition histologic coding can prevent this problem. This commentary examines the prevalence of this issue in published literature and highlights its potential impact on existing studies. Implementing rigorous coding practices and promoting their consistent use among researchers are crucial for ensuring valid and reliable findings in future pancreatic cancer studies using the UK Biobank.

Background: Personal health record (PHR) software has the potential of aiding with patient engagement and data collection in longitudinal research to better understand the long-term impact of treatments on patients with rare medical conditions. Neuroendocrine tumors (NETs) represent a rare condition with unique challenges related to symptom management, treatment tracking, and patient-provider communication. Objective: This study aimed to design, develop, and evaluate PHR software tailored for patients with NETs as part of a longitudinal research study. Our goal was to create a patient-centered PHR that supports both self-management and research data collection. Methods: This included activities spanning the entire development lifecycle from identifying user requirements through focus groups and surveys, iterative prototype refinement via cognitive walkthroughs, and usability testing of the functional PHR system. Feedback from patient advocacy organizations and clinical experts further informed PHR development. Results: The resulting PHR allows patients with NETs to access condition-specific information, track symptoms, monitor treatment regimens, and share data with health care providers. Patients valued the ability to visualize personal health trends and patterns over time, enhancing both self-management and communication with medical teams. Usability testing indicated high levels of patient satisfaction with the system’s functionality and design. Conclusions: The development of this PHR demonstrates the value of engaging patients in the design process to ensure that health technologies address real-world needs. Our approach provides a model for designing PHR systems for other rare conditions, highlighting the importance of patient-centered design in supporting both clinical care and longitudinal research.

PURPOSE: Electronic health records (EHRs) comprise a rich source of real-world data for cancer studies, but they often lack critical structured data elements such as diagnosis date and disease stage. Fortunately, such concepts are available from hospital cancer registries. We describe experiences from integrating cancer registry data with EHR and billing data in an interoperable data model across a multisite clinical research network. METHODS: After sites implemented cancer registry data into a tumor table compatible with the PCORnet Common Data Model (CDM), distributed queries were performed to assess quality issues. After remediation of quality issues, another query produced descriptive frequencies of cancer types and demographic characteristics. This included linked BMI. We also report two current use cases of the new resource. RESULTS: Eleven sites implemented the tumor table, yielding a resource with data for 572,902 tumors. Institutional and technical barriers were surmounted to accomplish this. Variations in racial and ethnic distributions across the sites were observed; the percent of tumors among Black patients ranged from <1% to 15% across sites, and the percent of tumors among Hispanic patients ranged from 1% to 46% across sites. Current use cases include a pragmatic prospective cohort study of a rare cancer and a retrospective cohort study leveraging body size and chemotherapy dosing. CONCLUSION: Integrating cancer registry data with the PCORnet CDM across multiple institutions creates a powerful resource for cancer studies. It provides a wider array of structured, cancer-relevant concepts, and it allows investigators to examine variability in those concepts across many treatment environments. Having the CDM tumor table in place enhances the impact of the network's effectiveness for real-world cancer research.

BACKGROUND Personal Health Record (PHR) software has the potential of aiding with patient engagement and data collection in longitudinal research to better understand the long-term impact of treatments on patients with rare medical conditions. OBJECTIVE We engaged with groups of patients with neuroendocrine tumors (NETs), a rare medical condition, to design PHR software to be used as part of a longitudinal research study with people with NETs. We sought to pursue a patient-centered approach to the design of the PHR. METHODS We describe our activities with multiple groups of patients that spanned the lifecycle from identifying requirements, iteratively developing a prototype, and refining a working application. Over 100 patients participated remotely in various activities including focus groups, surveys, cognitive walkthroughs, and usability testing. RESULTS Patients with NETs valued a PHR that allowed them to get information, track symptoms, and share symptom data with multiple medical providers. CONCLUSIONS Our experience suggests that the methods we used could be applied by other research teams with similar goals, and that it was valuable to engage patients throughout the process. CLINICALTRIAL Not applicable.

Objective We evaluate change in low back pain (LBP) intensity and interference as the mechanism by which chiropractic care affects other biopsychosocial factors in US active-duty military members. Design We conducted secondary, exploratory mediation analysis of pragmatic, multisite, clinical trial (NCT01692275) post results using natural effect modeling. Mediators were the 6-week values of Patient-Reported Outcomes Measurement Information System (PROMIS)-29 pain interference and intensity. Outcomes were 12-week values of other PROMIS-29 biopsychosocial subdomains. Models evaluated overall and individual factor contribution and were adjusted for baseline age, sex, LBP duration, LBP intensity and mediator and outcome values. Setting Three US military treatment facilities. Participants 750 US active-duty military members with LBP. Interventions Trial participants received 6 weeks of treatment with chiropractic care plus usual medical care or usual medical care alone. Results In multiple mediator models, pain interference and pain intensity explained much of the effect of chiropractic care on physical function (proportion mediated=0.77, 95% CI 0.43 to 2.0), fatigue (0.62, 95% CI 0.30 to 1.0), sleep disturbance (0.49, 95% CI 0.31 to 1.5) and social roles (0.81, 95% CI 0.50 to 2.0). Mental health was not evaluable due to the low prevalence of symptoms reported. The combined models of pain interference and pain intensity did not have a higher proportion mediated than the individual pain intensity models except for fatigue outcome models. Conclusion Pain intensity appeared to be the mechanism by which much of the change in biopsychosocial factors occurred and should be considered a key clinical indicator for improvement in biopsychosocial health when chiropractic care is added to usual medical care for US active-duty military members.