Using generative AI to investigate medical imagery models and datasets.

BACKGROUND: AI models have shown promise in performing many medical imaging tasks. However, our ability to explain what signals these models have learned is severely lacking. Explanations are needed in order to increase the trust of doctors in AI-based models, especially in domains where AI prediction capabilities surpass those of humans. Moreover, such explanations could enable novel scientific discovery by uncovering signals in the data that aren't yet known to experts. METHODS: In this paper, we present a workflow for generating hypotheses to understand which visual signals in images are correlated with a classification model's predictions for a given task. This approach leverages an automatic visual explanation algorithm followed by interdisciplinary expert review. We propose the following 4 steps: (i) Train a classifier to perform a given task to assess whether the imagery indeed contains signals relevant to the task; (ii) Train a StyleGAN-based image generator with an architecture that enables guidance by the classifier ("StylEx"); (iii) Automatically detect, extract, and visualize the top visual attributes that the classifier is sensitive towards. For visualization, we independently modify each of these attributes to generate counterfactual visualizations for a set of images (i.e., what the image would look like with the attribute increased or decreased); (iv) Formulate hypotheses for the underlying mechanisms, to stimulate future research. Specifically, present the discovered attributes and corresponding counterfactual visualizations to an interdisciplinary panel of experts so that hypotheses can account for social and structural determinants of health (e.g., whether the attributes correspond to known patho-physiological or socio-cultural phenomena, or could be novel discoveries). FINDINGS: To demonstrate the broad applicability of our approach, we present results on eight prediction tasks across three medical imaging modalities-retinal fundus photographs, external eye photographs, and chest radiographs. We showcase examples where many of the automatically-learned attributes clearly capture clinically known features (e.g., types of cataract, enlarged heart), and demonstrate automatically-learned confounders that arise from factors beyond physiological mechanisms (e.g., chest X-ray underexposure is correlated with the classifier predicting abnormality, and eye makeup is correlated with the classifier predicting low hemoglobin levels). We further show that our method reveals a number of physiologically plausible, previously-unknown attributes based on the literature (e.g., differences in the fundus associated with self-reported sex, which were previously unknown). INTERPRETATION: Our approach enables hypotheses generation via attribute visualizations and has the potential to enable researchers to better understand, improve their assessment, and extract new knowledge from AI-based models, as well as debug and design better datasets. Though not designed to infer causality, importantly, we highlight that attributes generated by our framework can capture phenomena beyond physiology or pathophysiology, reflecting the real world nature of healthcare delivery and socio-cultural factors, and hence interdisciplinary perspectives are critical in these investigations. Finally, we will release code to help researchers train their own StylEx models and analyze their predictive tasks of interest, and use the methodology presented in this paper for responsible interpretation of the revealed attributes. FUNDING: Google.

Care Partner Engagement in Secure Messaging Between Patients With Diabetes and Their Clinicians: Cohort Study.

BACKGROUND: Patient engagement with secure messaging (SM) via digital patient portals has been associated with improved diabetes outcomes, including increased patient satisfaction and better glycemic control. Yet, disparities in SM uptake exist among older patients and racial and ethnic underserved groups. Care partners (family members or friends) may provide a means for mitigating these disparities; however, it remains unclear whether and to what extent care partners might enhance SM use. OBJECTIVE: We aim to examine whether SM use differs among older patients with diabetes based on the involvement of care partner proxies. METHODS: This is a substudy of the ECLIPPSE (Employing Computational Linguistics to Improve Patient-Provider Secure Emails) project, a cohort study taking place in a large, fully integrated health care delivery system with an established digital patient portal serving over 4 million patients. Participants included patients with type 2 diabetes aged ≥50 years, newly registered on the patient portal, who sent ≥1 English-language message to their clinician between July 1, 2006, and December 31, 2015. Proxy SM was identified by having a registered proxy. To identify nonregistered proxies, a computational linguistics algorithm was applied to detect words and phrases more likely to appear in proxy messages compared to patient-authored messages. The primary outcome was the annual volume of secure messages (sent or received); secondary outcomes were the length of time to the first SM sent by patient or proxy and the number of annual SM exchanges (unique message topics generating ≥1 reply). RESULTS: The mean age of the cohort (N=7659) at this study's start was 61 (SD 7.16) years; 75% (n=5573) were married, 15% (n=1089) identified as Black, 10% (n=747) Chinese, 12% (n=905) Filipino, 13% (n=999) Latino, and 30% (n=2225) White. Further, 49% (n=3782) of patients used a proxy to some extent. Compared to nonproxy users, proxy users were older (P<.001), had lower educational attainment (P<.001), and had more comorbidities (P<.001). Adjusting for patient sociodemographic and clinical characteristics, proxy users had greater annual SM volume (20.7, 95% CI 20.2-21.2 vs 10.9, 95% CI 10.7-11.2; P<.001), shorter time to SM initiation (hazard ratio vs nonusers: 1.30, 95% CI 1.24-1.37; P<.001), and more annual SM exchanges (6.0, 95% CI 5.8-6.1 vs 2.9, 95% CI 2.9-3.0, P<.001). Differences in SM engagement by proxy status were similar across patient levels of education, and racial and ethnic groups. CONCLUSIONS: Among a cohort of older patients with diabetes, proxy SM involvement was independently associated with earlier initiation and increased intensity of messaging, although it did not appear to mitigate existing disparities in SM. These findings suggest care partners can enhance patient-clinician telecommunication in diabetes care. Future studies should examine the effect of care partners' SM involvement on diabetes-related quality of care and clinical outcomes.

Financial Strain and Medication Adherence among Diabetes Patients in an Integrated Health Care Delivery System: The Diabetes Study of Northern California (DISTANCE).

OBJECTIVE: To examine self-reported financial strain in relation to pharmacy utilization adherence data. DATA SOURCES/STUDY SETTING: Survey, administrative, and electronic medical data from Kaiser Permanente Northern California. STUDY DESIGN: Retrospective cohort design (2006, n = 7,773). DATA COLLECTION/EXTRACTION METHODS: We compared survey self-reports of general and medication-specific financial strain to three adherence outcomes from pharmacy records, specifying adjusted generalized linear regression models. PRINCIPAL FINDINGS: Eight percent and 9 percent reported general and medication-specific financial strain. In adjusted models, general strain was significantly associated with primary nonadherence (RR = 1.37; 95 percent CI: 1.04-1.81) and refilling late (RR = 1.34; 95 percent CI: 1.07-1.66); and medication-specific strain was associated with primary nonadherence (RR = 1.42, 95 percent CI: 1.09-1.84). CONCLUSIONS: Simple, minimally intrusive questions could be used to identify patients at risk of poor adherence due to financial barriers.

Use of the refill function through an online patient portal is associated with improved adherence to statins in an integrated health system.

BACKGROUND: Online patient portals are being widely implemented, but their impact on health behaviors are not well-studied. OBJECTIVE: To determine whether statin adherence improved after initiating use of the portal refill function. RESEARCH DESIGN: Observational cohort study within an integrated health care delivery system. SUBJECTS: Diabetic patients on statins who had registered for online portal access by 2010. A total of 8705 subjects initiated the online refill function use within the study window, including "exclusive" and "occasional" users (ie, requesting all vs. some refills online, respectively). Using risk-set sampling, we temporally matched 9055 reference group patients who never used online refills. MEASURES: We calculated statin adherence before and after refill function initiation, assessed as percent time without medications (nonadherence defined as a gap of >20%). Secondary outcome was dyslipidemia [low-density lipoprotein (LDL)≥ 100]. Difference-in-differences regression models estimated pre-post changes in nonadherence and dyslipidemia, comparing refill function users to the reference group and adjusting for age, sex, race/ethnicity, medications, frequency of portal use, and outpatient visits. RESULTS: In unadjusted examinations, nonadherence decreased only among patients initiating occasional or exclusive use of the refill function (26%-24% and 22%-15%, respectively). In adjusted models, nonadherence declined by an absolute 6% (95% confidence interval, 4%-7%) among exclusive users, without significant changes among occasional users. Similar LDL decreases were also seen among exclusive users. CONCLUSIONS: Compared with portal users who did not refill medications online, adherence to statin medications and LDL levels improved among diabetic patients who initiated and exclusively used the patient portal for refills, suggesting that wider adoption of online refills may improve adherence.