Elizabeth A. Bayliss

Background: Identifying clinically intended medication discontinuations at scale may help generate evidence to inform deprescribing. We developed an algorithmic approach to identifying such discontinuations, combining text strings applied to clinical documentation with medication order data for 5 different drug groups: oral hypoglycemics, statins, antihypertensives, bladder antimuscarinics, and antithrombotics. Design: The study population (n=1588) comprised individuals aged older than 65 with ≥90-day gaps in dispensing classified through manual review as having a clinically intended medication discontinuation or not. This gold-standard cohort was randomly divided into development and validation subsamples for each drug group. We developed text strings from clinical documentation reflecting clinical intent to discontinue a medication (or not). Text strings were usually simple [eg, “stop (medication)”], but required tailoring to drug groups [eg, temporarily “hold (antithrombotic)”]. Text strings queried clinical documentation and were supplemented with order discontinuation data, if available. We calculated sensitivity and specificity for identifying intended discontinuations for text alone, discrete data alone, and the full algorithm across validation samples. Results: Sensitivity and specificity for the full algorithm were 80% and 85% for oral hypoglycemics (n=467), 75% and 95% for statins (n=282), 82%, 75% for antihypertensives (n=599), 77% and 80% for bladder antimuscarinics (n=80), and 74% and 78% for antithrombotics (n=160). The full algorithm had higher sensitivity than either text or order data alone. Conclusions: Text-based approaches supplemented by medication order data can identify clinically intended medication discontinuations at scale with moderate specificity and sensitivity. This may reduce misclassification relative to using claims data to generate evidence about deprescribing.

Limited real-world evidence exists on the patterns of symptomatic treatment use for Alzheimer’s disease and related dementias (ADRD). Using data from

Background: Diabetes-related lower extremity complications, such as foot ulceration and amputation, are on the rise, currently affecting nearly 131 million people worldwide. Methods for early detection of individuals at high risk remain elusive. While data-driven diabetic polyneuropathy algorithms exist, high-performing, clinically useful tools to assess risk are needed to improve clinical care. Objective: This study aimed to develop an electronic medical record–based machine learning algorithm that would predict lower extremity complications. Methods: We conducted a retrospective longitudinal cohort study to predict the risk of lower extremity complications within 24 months of an initial diagnosis of diabetic polyneuropathy. From an initial cohort of 468,162 individuals with at least 1 diagnosis of diabetic polyneuropathy at one of 2 multispecialty health care systems (based in northern California and Colorado) between April 2012 and December 2016, we created an analytic cohort of 48,209 adults with continuous enrollment, who were newly diagnosed with no evidence of end-of-life care. The outcome was any lower extremity complication, including foot ulceration, osteomyelitis, gangrene, or lower extremity amputation. We randomly split the data into training (38,569/48209; 80%) and testing (9,640/48209; 20%) datasets. In the training dataset, we used super Learner (SL), an ensemble learning method that employs cross-validation and combines multiple candidate risk predictors, into a single risk predictor. We evaluated the performance of the SL risk predictor in the testing dataset using the receiver operating characteristic curve and a calibration plot. Results: Of the 48,209 individuals in the cohort, 2327 developed a lower extremity complication during follow-up. The SL risk estimator exhibited good discrimination (AUC=0.845, 95% CI 0.826-0.863) and calibration. A modified version of our SL algorithm, simplified to facilitate real-world adoption, had only slightly reduced discrimination (AUC=0.817, 95%CI 0.797-0.837). The modified version slightly outperformed the naïve logistic regression model (AUC=0.804, 95% CI 0.783-0.825) in terms of precision gained relative to the frequency of alerts and number of patients that needed to be evaluated. Conclusions: We have built a machine learning–based risk estimator with the potential to improve clinical detection of diabetic patients at high risk for lower extremity complications at the time of an initial diabetic polyneuropathy diagnosis. The algorithm exhibited good discriminant validity and calibration using only data from the electronic medical record. Additional research will be needed to identify optimal contexts and strategies for maximizing algorithmic fairness in both interpretation and deployment.

Background: Systemically identifying caregivers in the electronic health record (EHR) is a critical step for delivering patient-centered care, enhancing care coordination, and advancing research and population health efforts in caregiving. Despite EHRs being effective in identifying patients through standardized data fields like demographics, laboratory results, medications, and diagnoses, identifying caregivers through the EHR is challenging in the absence of specific caregiver fields. Objective: Recognizing the complexity of identifying caregiving networks of people living with dementia, this study aims to systematically capture caregiver information by combining EHR structured fields, unstructured notes, and free text. Methods: Among a cohort of people living with dementia aged 60 years and older from Kaiser Permanente Colorado, caregiver names were identified by combining structured patient contact fields, that is, known contacts, with name-matching and natural language processing techniques of unstructured notes and patient portal messages containing caregiver terms. Results: Among the cohort of 789 people living with dementia, 95% (n=749) had at least 1 caregiver name listed in structured fields (mean 2.1 SD 1.1). Over 95% of the cohort had caregiver terms mentioned near a known contact name in unstructured encounter notes, with 35% having a full name match in unstructured patient portal messages. The natural language processing model identified 7556 “new” names in the unstructured EHR text containing caregiver terms among 99% of the cohort with high accuracy and reliability (F1-score=.85; precision=.89; recall=.82). Overall, 87% of the cohort had a new name identified ≥2 near a caregiver term in their notes and portal messages. Conclusions: Patterns in caregiver-related information were distributed across structured and unstructured EHR fields, emphasizing the importance of integrating both data sources for a comprehensive understanding of caregiving networks. A framework was developed to systematically identify potential caregivers across caregiving networks using structured and unstructured EHR data. This approach has the potential to improve health services for people living with dementia and their caregivers.

Purpose: Accurately identifying medication discontinuations at scale is important for developing evidence about deprescribing. Gaps in dispensing often serve as proxies for discontinuation but are imprecise. We categorize reasons for gaps in dispensing to inform data‐based methods to accurately identify medication discontinuations.Methods: Using pharmacy dispensing data, we purposively sampled from a population of adults age 65+ with 2+ chronic conditions who experienced a 90‐day gap in dispensing—with and without subsequent fills—of oral diabetes drugs, statins, proton pump inhibitors, drugs with anticholinergic properties, anticoagulants and antiplatelet drugs, or antihypertensives. We reviewed clinical documentation (e.g., visit notes, communications, medication orders) from last dispensing through the 90‐day gap plus 120 days to classify dispensing gaps as true discontinuations (clinically intended) or non‐discontinuations (no evidence of intent to discontinue), and then into subcategories. Medications with no documented explanation for the gap in dispensing and continued listing on the patient's medication list were classified as non‐discontinuations.Results: Of N = 1906 records reviewed, there were 1068 (56%) true discontinuations and 838 (44%) non‐discontinuations. Subcategories within true discontinuations included provider intent to discontinue, provider substitutions, intentional stops followed by restarts, and agreeing with a colleague's or patient's decision to discontinue. Non‐discontinuations included documented low adherence, changes in dose, changes in pharmacy formulary, and changes in drug formulation. Proportions of drugs in categories and subcategories varied by medication group.Conclusion: Using gaps in dispensing as proxy measures for medication discontinuation may introduce bias through misclassification, and varied reasons for discontinuation may complicate causal interpretations.

Background: Family and unpaid caregivers play a crucial role in supporting people living with dementia; yet, they are not systematically identified and documented by health systems. Objective: The aims of the study are to determine the extent to which caregivers are currently identified and documented in the electronic health record (EHR) and to elicit the perspectives of caregivers and clinical staff on how to best identify, engage, and support caregivers of people living with dementia through the EHR. Methods: People with dementia were identified based on International Classification of Diseases, Tenth Revision (ICD-10) codes or dementia medications in the EHR. A chart review of people with dementia characterized how caregiver information was documented and whether caregivers had shared access to the patient portal. Caregivers of eligible people with dementia were then recruited through mailed letters and follow-up calls to the homes of people with dementia. We conducted semistructured interviews with caregivers, clinicians, and staff involved in the care of people with dementia within 2 health systems in Maryland and Colorado. Transcripts were analyzed using a mixed inductive and deductive approach. Results: Caregivers of people with dementia (N=22) were usually identified in the “contact information” or “patient contacts” tab (n=20, 91%) by their name and relation to the people with dementia; this tab did not specify the caregiver’s role. Caregivers were also mentioned, and their roles were described to a varying degree in clinical notes (n=21, 96%). Of the 22 caregivers interviewed, the majority (n=17, 77%) reported that the people with dementia had additional caregivers. The presence of multiple caregivers could be gleaned from most charts (n=16, 73%); however, this information was not captured systematically, and caregivers’ individual contributions were not explicitly recorded. Interviews with 22 caregivers and 16 clinical staff revealed two major themes: (1) caregiving arrangements are complex and not systematically captured or easy to locate in the EHR and (2) health systems should develop standardized processes to obtain and document caregiver information in the EHR. Conclusions: This exploratory chart review and qualitative interview study found that people with dementia frequently have multiple caregivers, whose roles and needs are captured inconsistently in the EHR. To address this concern, caregivers and clinical staff suggested that health systems should develop and test workflows to identify caregivers, assess their needs at multiple touchpoints, and record their information in extractable EHR fields.

Generating real‐world evidence about the effect of medication discontinuation or dose reduction on outcomes, such as reduction of adverse drug effects (ADE; intended benefit) and occurrence of adverse drug withdrawal events (ADWE; unintended harm), is crucial to informing deprescribing decisions. Determining the causal effects of deprescribing is difficult for many reasons, including lack of randomization in real‐world study designs and other design and measurement issues that pose threats to internal validity. The inherent challenge is how to identify the effects, both intended benefits and unintended harms, of a new medication stoppage or reduction when implemented in patients with many potential clinical and social risks that may influence the likelihood of deprescribing as well as outcomes. We discuss methodological issues of estimating the effect of medication discontinuation or reduction on risk of ADEs and ADWEs considering: (1) sampling study populations of sufficient size with the potential to demonstrate clinically meaningful and quantifiable outcomes, (2) accurate and appropriately timed measurement of covariates, outcomes, and discontinuation, and (3) statistical approaches to managing confounding and other biases inherent in long‐term medication use by individuals with multiple morbidities. Designing rigorous deprescribing studies that address internal validity threats will support evidence generation by improving the ability to assess benefits and harms when the exposure of interest is the absence of a medication. Iterative learnings about data quality, variable definition, variable measurement, and exposure‐outcome associations will inform strategies to improve the causal inferences possible in real‐world deprescribing studies.

BACKGROUND Systemically identifying caregivers in the electronic health record (EHR) is a critical step for delivering patient-centered care, enhancing care coordination, and advancing research and population health efforts in caregiving. Despite EHRs being effective in identifying patients through standardized data fields like demographics, laboratory results, medications, and diagnoses, identifying caregivers through the EHR is challenging in the absence of specific caregiver fields. OBJECTIVE Recognizing the complexity of identifying caregiving networks of people living with dementia, this study aims to systematically capture caregiver information by combining EHR structured fields, unstructured notes, and free text. METHODS Among a cohort of people living with dementia aged 60 years and older from Kaiser Permanente Colorado, caregiver names were identified by combining structured patient contact fields, that is, known contacts, with name-matching and natural language processing techniques of unstructured notes and patient portal messages containing caregiver terms. RESULTS Among the cohort of 789 people living with dementia, 95% (n=749) had at least 1 caregiver name listed in structured fields (mean 2.1 SD 1.1). Over 95% of the cohort had caregiver terms mentioned near a known contact name in unstructured encounter notes, with 35% having a full name match in unstructured patient portal messages. The natural language processing model identified 7556 “new” names in the unstructured EHR text containing caregiver terms among 99% of the cohort with high accuracy and reliability (F1-score=.85; precision=.89; recall=.82). Overall, 87% of the cohort had a new name identified ≥2 near a caregiver term in their notes and portal messages. CONCLUSIONS Patterns in caregiver-related information were distributed across structured and unstructured EHR fields, emphasizing the importance of integrating both data sources for a comprehensive understanding of caregiving networks. A framework was developed to systematically identify potential caregivers across caregiving networks using structured and unstructured EHR data. This approach has the potential to improve health services for people living with dementia and their caregivers.