Background: Machine learning is increasingly used to predict healthcare outcomes, including cost, utilization, and quality.Objective: We provide a high-level overview of machine learning for healthcare outcomes researchers and decision makers.Methods: We introduce key concepts for understanding the application of machine learning methods to healthcare outcomes research. We first describe current standards to rigorously learn an estimator, which is an algorithm developed through machine learning to predict a particular outcome. We include steps for data preparation, estimator family selection, parameter learning, regularization, and evaluation. We then compare 3 of the most common machine learning methods: (1) decision tree methods that can be useful for identifying how different subpopulations experience different risks for an outcome; (2) deep learning methods that can identify complex nonlinear patterns or interactions between variables predictive of an outcome; and (3) ensemble methods that can improve predictive performance by combining multiple machine learning methods. Results:We demonstrate the application of common machine methods to a simulated insurance claims dataset. We specifically include statistical code in R and Python for the development and evaluation of estimators for predicting which patients are at heightened risk for hospitalization from ambulatory care-sensitive conditions.Conclusions: Outcomes researchers should be aware of key standards for rigorously evaluating an estimator developed through machine learning approaches. Although multiple methods use machine learning concepts, different approaches are best suited for different research problems.

This secondary analysis of the Systolic Blood Pressure Intervention Trial uses a random forest–based analysis to assess whether clinically important heterogeneity exists in the risk of harm associated with intensive blood pressure treatment among adults with hypertension who smoke.

ObjectiveTo (1) examine the burden of multiple chronic conditions (MCC) in an urban health system, and (2) propose a methodology to identify subpopulations of interest based on diagnosis groups and costs.DesignRetrospective cross-sectional study.SettingMount Sinai Health System, set in all five boroughs of New York City, USA.Participants192 085 adult (18+) plan members of capitated Medicaid contracts between the Healthfirst managed care organisation and the Mount Sinai Health System in the years 2012 to 2014.MethodsWe classified adults as having 0, 1, 2, 3, 4 or 5+ chronic conditions from a list of 69 chronic conditions. After summarising the demographics, geography and prevalence of MCC within this population, we then described groups of patients (segments) using a novel methodology: we combinatorially defined 18 768 potential segments of patients by a pair of chronic conditions, a sex and an age group, and then ranked segments by (1) frequency, (2) cost and (3) ratios of observed to expected frequencies of co-occurring chronic conditions. We then compiled pairs of conditions that occur more frequently together than otherwise expected.Results61.5% of the study population suffers from two or more chronic conditions. The most frequent dyad was hypertension and hyperlipidaemia (19%) and the most frequent triad was diabetes, hypertension and hyperlipidaemia (10%). Women aged 50 to 65 with hypertension and hyperlipidaemia were the leading cost segment in the study population. Costs and prevalence of MCC increase with number of conditions and age. The disease dyads associated with the largest observed/expected ratios were pulmonary disease and myocardial infarction. Inter-borough range MCC prevalence was 16%.ConclusionsIn this low-income, urban population, MCC is more prevalent (61%) than nationally (42%), motivating further research and intervention in this population. By identifying potential target populations in an interpretable manner, this segmenting methodology has utility for health services analysts.

Objective Community health systems operating in remote areas require accurate information about where people live to efficiently provide services across large regions. We sought to determine whether a machine learning analyses of satellite imagery can be used to map remote communities to facilitate service delivery and planning. Materials and Methods We developed a method for mapping communities using a deep learning approach that excels at detecting objects within images. We trained an algorithm to detect individual buildings, then examined building clusters to identify groupings suggestive of communities. The approach was validated in southeastern Liberia, by comparing algorithmically generated results with community location data collected manually by enumerators and community health workers. Results The deep learning approach achieved 86.47% positive predictive value and 79.49% sensitivity with respect to individual building detection. The approach identified 75.67% (n = 451) of communities registered through the community enumeration process, and identified an additional 167 potential communities not previously registered. Several instances of false positives and false negatives were identified. Discussion Analysis of satellite images is a promising solution for mapping remote communities rapidly, and with relatively low costs. Further research is needed to determine whether the communities identified algorithmically, but not registered in the manual enumeration process, are currently inhabited. Conclusions To our knowledge, this study represents the first effort to apply image recognition algorithms to rural healthcare delivery. Results suggest that these methods have the potential to enhance community health worker scale-up efforts in underserved remote communities.

Precision medicine research designed to reduce health disparities often involves studying multi-level datasets to understand how diseases manifest disproportionately in one group over another, and how scarce health care resources can be directed precisely to those most at risk for disease. In this article, we provide a structured tutorial for medical and public health research­ers on the application of machine learning methods to conduct precision medicine research designed to reduce health dispari­ties. We review key terms and concepts for understanding machine learning papers, including supervised and unsupervised learning, regularization, cross-validation, bagging, and boosting. Metrics are reviewed for evaluating machine learners and major families of learning approaches, including tree-based learning, deep learning, and ensemble learning. We highlight the advan­tages and disadvantages of different learning approaches, describe strategies for interpret­ing “black box” models, and demonstrate the application of common methods in an example dataset with open-source statistical code in R.Ethn Dis. 2020;30(Suppl 1):217-228; doi:10.18865/ed.30.S1.217

No abstract

No abstract

To measure the deforestation reduced by a policy, we need to compare deforestation rates under a policy with deforestation rates in the absence of policy. Unfortunately the deforestation rate in the absence of a policy, or reference rate, is ex ante difficult to forecast and ex post impossible to observe. This means that reference rates will be set with error and we will not know how large the error will be. The challenging nature of setting reference rates is reflected in the number of proposals for reference rate design. In this paper I show how these proposals ignore forecast error. As a consequence, these proposals have basic structural weaknesses that increase the costs of reduced deforestation policy. I propose that a criteria for reference rates is to minimise the cost of forecast error. These ideas are illustrated with a cross country dataset on agricultural expansion. I show that the best forecasting model differs by country and that a country's best forecasting model can be very simple.