 On March 10th, the first case of COVID-19 was confirmed in the state of Michigan. Throughout the remainder of the month, the total number of cases quickly rose. Though Michigan Medicine started with just a handful of cases, without intervention, it expected the number of admitted patients with COVID-19 to increase exponentially in the coming days and weeks. This exponential increase combined with a potential shortage of resources led to tough questions. Here at Michigan, hospital leadership began to explore converting an indoor track into a field hospital to meet the demand. Around this time, members of the AI Lab partnered with members of Michigan Medicine to help develop clinician decision support tools. Specifically, tools that could help identify patients at greatest risk of complications or adverse outcomes from the disease, for example those who might need respiratory support, including a mechanical ventilator, cardiovascular support, and those who would likely die in the hospital from the disease. Such a predictive model could help hospital manage a surge in COVID-19 patients and anticipate needs in advance. Given that along with the rest of the world, we knew little about this disease and potential risk factors, we took a data-centric approach to the problem, leveraging the contents of the electronic health record to learn and validate models for predicting adverse outcomes. Our system, dubbed M-Cures, short for Michigan COVID-19 Utilization and Risk Evaluation System, uses data collected in the EHR, like medications and vital sign measurements, and estimates the patient's risk of experiencing an adverse outcome. This allows clinicians to rank patients from low risk to high risk. M-Cures repeatedly reassesses and re-ranks patients over time throughout the hospitalization as new information becomes available. From a machine learning for healthcare perspective, we encountered numerous challenges, including the fact that though cases were increasing rapidly, there was relatively little training data. To tackle this issue, we leveraged data from a large retrospective cohort of individuals presenting in the emergency department with respiratory symptoms. We held out individuals with COVID-19 for evaluation purposes, but again, due to the limited number of cases and the accelerated deployment timeline, we had to find alternate ways to scrutinize the model to limit issues in real-time deployment. Finally, though we started with thousands of variables extracted from the EHR in an effort to produce a model that could serve not only Michigan medicine, but potentially other hospitals across the globe, we aimed to reduce the model to a subset of readily available data in the EHR that could easily transfer across institutions and populations. We evaluated the model's predictions on a held-out test cohort of over 200 patients. Working closely with clinicians in the hospital, we focused on two potential use cases for the model. The first aims to identify individuals who will experience an adverse outcome early during an admission. Nearly 30% of the test cohort met this outcome, and the M-CURES model achieved good discriminative performance, correctly identifying nearly half of the high-risk cases while limiting the number of false positives, significantly outperforming a proprietary deterioration index integrated in the EHR. We also explored the ability of the model to identify low-risk individuals, or patients who are unlikely to suffer an adverse outcome in the remainder of the admission. These are patients that could be safely transferred to a field hospital. In terms of the outcomes we explored, 78% of our test cohort were deemed low-risk. M-CURES flagged nearly 20% of patients as low-risk, while incurring zero false positives. This did significantly better or flag more low-risk patients than the proprietary model. Around the same time that we were working on these clinical decision support tools, the governor implemented a stay-at-home order, and fortunately the curve flattened. Michigan Medicine did not have to open the field hospital, and clinicians did not have to turn to using M-CURES. And while we do not know what the future holds, like many states across the country, Michigan could experience another surge in cases. If other measures are effective and the number of cases does not surge, M-CURES could still help us manage admitted patients alongside non-COVID patients, especially in the presence of a bad flu season. Finally, though the model was evaluated on COVID-19 patients, it could be extended as an early warning system for the general inpatient population, where it performed similarly. We are now working with the Clinical Intelligence Committee at Michigan Medicine to integrate the model into the EHR. Though we were scattered across the state working from our living rooms, studies, and dining tables, we came together as a team. Overall, a process that typically takes years was achieved in the matter of weeks. Given the accelerated timeline, we learned a lot that will hopefully inform future models and ensure we're more prepared next time.