Faculty Bibliography

Testing multiple treatments for heterogeneous (varying) effectiveness with respect to many underlying risk factors requires many pairwise tests; we would like to instead automatically discover and visualize patient archetypes and predictors of treatment effectiveness using multitask machine learning. In this paper, we present a method to estimate these heterogeneous treatment effects with an interpretable hierarchical framework that uses additive models to visualize expected treatment benefits as a function of patient factors (identifying personalized treatment benefits) and concurrent treatments (identifying combinatorial treatment benefits). This method achieves state-of-the-art predictive power for COVID-19 in-hospital mortality and interpretable identification of heterogeneous treatment benefits. We first validate this method on the large public MIMIC-IV dataset of ICU patients to test recovery of heterogeneous treatment effects. Next we apply this method to a proprietary dataset of over 3000 patients hospitalized for COVID-19, and find evidence of heterogeneous treatment effectiveness predicted largely by indicators of inflammation and thrombosis risk: patients with few indicators of thrombosis risk benefit most from treatments against inflammation, while patients with few indicators of inflammation risk benefit most from treatments against thrombosis. This approach provides an automated methodology to discover heterogeneous and individualized effectiveness of treatments.

Machine learning applications promise to augment clinical capabilities and at least 64 models have already been approved by the US Food and Drug Administration. These tools are developed, shared, and used in an environment in which regulations and market forces remain immature. An important consideration when evaluating this environment is the introduction of open-source solutions in which innovations are freely shared; such solutions have long been a facet of digital culture. We discuss the feasibility and implications of open-source machine learning in a health care infrastructure built upon proprietary information. The decreased cost of development as compared to drugs and devices, a longstanding culture of open-source products in other industries, and the beginnings of machine learning-friendly regulatory pathways together allow for the development and deployment of open-source machine learning models. Such tools have distinct advantages including enhanced product integrity, customizability, and lower cost, leading to increased access. However, significant questions regarding engineering concerns about implementation infrastructure and model safety, a lack of incentives from intellectual property protection, and nebulous liability rules significantly complicate the ability to develop such open-source models. Ultimately, the reconciliation of open-source machine learning and the proprietary information-driven health care environment requires that policymakers, regulators, and health care organizations actively craft a conducive market in which innovative developers will continue to both work and collaborate.

Background: Risk calculators to predict perioperative major adverse cardiovascular and cerebrovascular events (MACCE) often rely on logistic regression (LR) analysis. Automated machine learning (AutoML) processes regularly outperform regular machine learning (ML) and LR methods for predictive accuracy. Commercial AutoML systems have not yet been applied to predict perioperative MACCE after non-cardiac and cardiac surgeries.
Method(s): We used a commercial AutoML system, Google AutoML Tables (GAMLT), to predict perioperative MACCE in the 2019 ACS NSQIP database. MACCE was defined as death, myocardial infarction, cardiac arrest, or stroke. Default AutoML settings were used, with 80% of cases randomly selected for training, 10% for validation, and 10% for testing. Global feature importance was determined through the Shapley method. Two models were generated: Model 1 included 81 pre-operative features; Model 2 included the top 21 features from Model 1 and was independently validated with 2016-2018 NSQIP data.
Result(s): Model 1 yielded an area under the receiver operating characteristic (AUROC) of 0.934 in the 2019 ACS NSQIP dataset. Model 2, an ensemble of 25 feedforward neural net models, yielded an AUROC of 0.914-0.920 for MACCE in 2016-2019 (Figure).
Conclusion(s): Compared to existing risk calculators, GAMLT-derived models offered novel feature detection and comparable predictive performance for MACCE. AutoML analyses should be considered for risk estimation of perioperative MACCE. [Formula presented]
Copyright

BACKGROUND:Residents and fellows receive little feedback on their clinical reasoning documentation. Barriers include lack of a shared mental model and variability in the reliability and validity of existing assessment tools. Of the existing tools, the IDEA assessment tool includes a robust assessment of clinical reasoning documentation focusing on four elements (interpretive summary, differential diagnosis, explanation of reasoning for lead and alternative diagnoses) but lacks descriptive anchors threatening its reliability. OBJECTIVE:Our goal was to develop a valid and reliable assessment tool for clinical reasoning documentation building off the IDEA assessment tool. DESIGN, PARTICIPANTS, AND MAIN MEASURES/UNASSIGNED:The Revised-IDEA assessment tool was developed by four clinician educators through iterative review of admission notes written by medicine residents and fellows and subsequently piloted with additional faculty to ensure response process validity. A random sample of 252 notes from July 2014 to June 2017 written by 30 trainees across several chief complaints was rated. Three raters rated 20% of the notes to demonstrate internal structure validity. A quality cut-off score was determined using Hofstee standard setting. KEY RESULTS/RESULTS:The Revised-IDEA assessment tool includes the same four domains as the IDEA assessment tool with more detailed descriptive prompts, new Likert scale anchors, and a score range of 0-10. Intraclass correlation was high for the notes rated by three raters, 0.84 (95% CI 0.74-0.90). Scores ≥6 were determined to demonstrate high-quality clinical reasoning documentation. Only 53% of notes (134/252) were high-quality. CONCLUSIONS:The Revised-IDEA assessment tool is reliable and easy to use for feedback on clinical reasoning documentation in resident and fellow admission notes with descriptive anchors that facilitate a shared mental model for feedback.

Predictive models may be particularly beneficial to clinicians when they face uncertainty and seek to develop a mental model of disease progression, but we know little about the post-implementation effects of predictive models on clinicians' experience of their work. Combining survey and interview methods, we found that providers using a predictive algorithm reported being significantly less uncertain and better able to anticipate, plan and prepare for patient discharge than non-users. The tool helped hospitalists form and develop confidence in their mental models of a novel disease (Covid-19). Yet providers' attention to the predictive tool declined as their confidence in their own mental models grew. Predictive algorithms that not only offer data but also provide feedback on decisions, thus supporting providers' motivation for continuous learning, hold promise for more sustained provider attention and cognition augmentation.

OBJECTIVE:Heightened inflammation, dysregulated immunity, and thrombotic events are characteristic of hospitalized COVID-19 patients. Given that platelets are key regulators of thrombosis, inflammation, and immunity they represent prime candidates as mediators of COVID-19-associated pathogenesis. The objective of this study was to understand the contribution of severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) to the platelet phenotype via phenotypic (activation, aggregation) and transcriptomic characterization. APPROACH AND RESULTS/UNASSIGNED:In a cohort of 3915 hospitalized COVID-19 patients, we analyzed blood platelet indices collected at hospital admission. Following adjustment for demographics, clinical risk factors, medication, and biomarkers of inflammation and thrombosis, we find platelet count, size, and immaturity are associated with increased critical illness and all-cause mortality. Bone marrow, lung tissue, and blood from COVID-19 patients revealed the presence of SARS-CoV-2 virions in megakaryocytes and platelets. Characterization of COVID-19 platelets found them to be hyperreactive (increased aggregation, and expression of P-selectin and CD40) and to have a distinct transcriptomic profile characteristic of prothrombotic large and immature platelets. In vitro mechanistic studies highlight that the interaction of SARS-CoV-2 with megakaryocytes alters the platelet transcriptome, and its effects are distinct from the coronavirus responsible for the common cold (CoV-OC43). CONCLUSIONS:Platelet count, size, and maturity associate with increased critical illness and all-cause mortality among hospitalized COVID-19 patients. Profiling tissues and blood from COVID-19 patients revealed that SARS-CoV-2 virions enter megakaryocytes and platelets and associate with alterations to the platelet transcriptome and activation profile.

BACKGROUND:It is estimated that up to 40% of Alzheimer's Disease and Related Dementias cases can be prevented or delayed by addressing modifiable factors including those that influence vascular risk (hypertension, obesity, smoking, physical activity, diabetes). Prevention may be particularly important in the vascular dementia subtypes. Despite the supporting evidence, the rates of medical therapy to reduce vascular risk are not well described. METHOD/METHODS:We assessed the utilization of statins, aspirin, and blood pressure (BP) medications in adults age ≥65 years cared for at NYU Langone Health, as recorded in the electronic health record. We included two cohorts: cohort 1 included patients who were diagnosed with vascular dementia (VaD) at NYU Langone Barlow Center for Memory Evaluation between January 1, 2015 and June 24, 2019. Cohort 2 extended the inclusion to seniors with VD diagnosis by any NYU Langone physician. Definitions for vascular dementia, the covariates assessed, and medications that we included in each category are shown in Tables 1-3. RESULT/RESULTS:We included 419 and 3745 patients in cohort 1 and cohort 2, respectively. Table 4 shows the characteristics and medication adherence in cohorts 1 and 2. In cohort 1, the prescription rates for statins, aspirin, and BP medications were 66%, 66%, 70%. In cohort 2, the rates for statin, aspirin, and BP medications were 56%, 46%, and 65%, respectively. The differences between prescription rates in cohort 1 and 2 for the three medication groups were statistically significant (p<0.05). CONCLUSION/CONCLUSIONS:Our analysis of the utilization of cardiovascular medications among patients with vascular dementia illuminates potential gaps both among patients who receive care at specialty clinics, as well as the overall population with vascular dementia. The rates of medication utilization are higher for patients under the care of cognitive neurologists. Electronic health records can help identify large cohorts of patients who may benefit from improved access to preventative measures including cardiovascular medications.

PURPOSE:Residency programs face overwhelming numbers of residency applications, limiting holistic review. Artificial intelligence techniques have been proposed to address this challenge but have not been created. Here, a multidisciplinary team sought to develop and validate a machine learning (ML)-based decision support tool (DST) for residency applicant screening and review. METHOD:Categorical applicant data from the 2018, 2019, and 2020 residency application cycles (n = 8,243 applicants) at one large internal medicine residency program were downloaded from the Electronic Residency Application Service and linked to the outcome measure: interview invitation by human reviewers (n = 1,235 invites). An ML model using gradient boosting was designed using training data (80% of applicants) with over 60 applicant features (e.g., demographics, experiences, academic metrics). Model performance was validated on held-out data (20% of applicants). Sensitivity analysis was conducted without United States Medical Licensing Examination (USMLE) scores. An interactive DST incorporating the ML model was designed and deployed that provided applicant- and cohort-level visualizations. RESULTS:The ML model areas under the receiver operating characteristic and precision recall curves were 0.95 and 0.76, respectively; these changed to 0.94 and 0.72, respectively, with removal of USMLE scores. Applicants' medical school information was an important driver of predictions-which had face validity based on the local selection process-but numerous predictors contributed. Program directors used the DST in the 2021 application cycle to select 20 applicants for interview that had been initially screened out during human review. CONCLUSIONS:The authors developed and validated an ML algorithm for predicting residency interview offers from numerous application elements with high performance-even when USMLE scores were removed. Model deployment in a DST highlighted its potential for screening candidates and helped quantify and mitigate biases existing in the selection process. Further work will incorporate unstructured textual data through natural language processing methods.

OBJECTIVES/OBJECTIVE:Predictive studies play important roles in the development of models informing care for patients with COVID-19. Our concern is that studies producing ill-performing models may lead to inappropriate clinical decision-making. Thus, our objective is to summarise and characterise performance of prognostic models for COVID-19 on external data. METHODS:We performed a validation of parsimonious prognostic models for patients with COVID-19 from a literature search for published and preprint articles. Ten models meeting inclusion criteria were either (a) externally validated with our data against the model variables and weights or (b) rebuilt using original features if no weights were provided. Nine studies had internally or externally validated models on cohorts of between 18 and 320 inpatients with COVID-19. One model used cross-validation. Our external validation cohort consisted of 4444 patients with COVID-19 hospitalised between 1 March and 27 May 2020. RESULTS:Most models failed validation when applied to our institution's data. Included studies reported an average validation area under the receiver-operator curve (AUROC) of 0.828. Models applied with reported features averaged an AUROC of 0.66 when validated on our data. Models rebuilt with the same features averaged an AUROC of 0.755 when validated on our data. In both cases, models did not validate against their studies' reported AUROC values. DISCUSSION/CONCLUSIONS:Published and preprint prognostic models for patients infected with COVID-19 performed substantially worse when applied to external data. Further inquiry is required to elucidate mechanisms underlying performance deviations. CONCLUSIONS:Clinicians should employ caution when applying models for clinical prediction without careful validation on local data.