Agenda

Blood, platelets and other transferable fluids are critical for patient health.  At PAW-2020 we described OneBlood’s use of analytics to optimize blood donor recruitment, to forecast hospital needs and to manage the blood supply chain during the Covid pandemic.  Now we provide an update with a focus on platelets.  We built and deployed three predictive models.  Marketing campaigns use these models, and dashboards enable campaign tracking and iterative improvements.  Inventory monitoring and hospital demand forecasting are the remaining solution components.  With all of these components working together, we have dramatically increased the number of platelet donors, stabilized inventory to match demand, and dramatically increased platelet availability in Florida hospitals.

In January 2020, just after the first case of COVID-19 was discovered in the US, NYU Computer Science Prof. Anasse Bari led with infectious disease medical doctor Prof. Megan Coffee a multidisciplinary team of AI and medicine experts both in the US and China to develop the first COVID-19 Clinical Severity Predictive Tool. The tool aimed to help medical doctors triage and provide care effectively during the incoming surges of cases by using algorithms that can predict which mildly ill patients were likely to become severely ill. In July 2021, the team developed another tool named COVID-19 Early-alerts Signals built on a digital epidemiology framework that analyzes alternative data sources to discover predictors of the pandemic curve, which could supplement traditional predictive models and inform early warning systems and public health policies. The research finds that online google searches can predict major regional increases and decreases in COVID-19 cases.  After the vaccine rollout, Prof. Bari and Coffee led a team that developed a Vaccine Hesitancy Analytics Tool which is a real-time big data analytics cloud application to track misinformation and extract themes and topics related to vaccine hesitancy. The platform was based on natural language processing and sentiment analysis predictive algorithms. The tool was deployed using Amazon Web Services. 

 
In this talk Prof. Bari will outline the experimental research results from the three tools his team developed: (1) COVID-19 Clinical Severity Predictor, (2) Pandemics Early-alert Signals Tool based on alternative data, and (3) Vaccine Hesitancy Analytics Tool. This talk will also highlight the analytics lessons learned and how we can better prepare for future pandemics using predictive analytics and algorithms.
 
* Prof. Anasse Bari led these projects and teams with medical doctor Prof. Megan Coffee, Dr. Matthias Heymann and other researchers from the NYU Courant Institute of Mathematical Sciences, the NYU Computer Science Department and the NYU Grossman School of Medicine. 

Building and deploying predictive models for the COVID-19 pandemic was challenging and most of the models have not performed as well as hoped.  I cover five lessons learned from analyzing data by the Pandemic Response Commons, a not-for-profit that collects, analyzes and shares COVID-19 related data in the Chicago region.  I also look at the challenges of understanding COVID-19 health disparities and present the results of models showing the unequal impact of COVID-19 on different populations.  We conclude by discussing how regions can prepare for the future by putting in place persistent infrastructure for regional data collection, analysis and sharing.

Timely capacity planning of cardiothoracic ICU relies on early doctor’s recommendations on the type of bed a patient will need after surgery. A predictive model might be used to potentially help doctors make more accurate recommendations. Using the data of around 1500 cardiothoracic surgeries we built a gradient boosting model which predicts whether a patient will need a PACU (i.e. fast track trajectory) or an ICU bed with AUC=0.8 (precision=0.59, recall=0.84). In a hybrid scenario where the recommendation for PACU trajectory is made combining doctor’s recommendations and model’s predictions, the number of patients misclassified to PACU would be 1-in-10, which would be a significant improvement over the current 1-in-5.

It is difficult to estimate whether any individual patient will refill their prescription, even if we provide them a scheduled reminder. Inferring whether properties of that reminder are important, such as when it is delivered, is more difficult. Bayesian multi-factor models (also known as hierarchical or random effects models) offer a robust way to model this problem, accounting for differences across patients and explicitly accounting for model uncertainties. Fitting robust Bayesian models can be difficult in practice, but in the last few years work has been devoted to "Bayesian workflows," which offer principled approaches to building such models. In this talk, we will use one recent project, in which we built a model to understand the impact of reminder-refill timing, to discuss key elements of the workflow and how Bayesian techniques provide us with robust models that are able to quantify the uncertainties inherent in statistical inference.

Long-term health consequences of COVID-19 are symptoms that continue weeks or months after first diagnoses. Symptoms span respiratory, neurological, psychological, and cardiac problems and range from mild to debilitating.  Little is known about the risk factors contributing to long COVID, whether vaccines play a role or the best treatment options. UnitedHealth Group data represents millions of COVID-19 patients – some fully recovered and others that suffer from continued health consequences. Machine learning provides the opportunity to characterize these risk factors and predict probability that future disease will occur.

A key to meaningfully and sustainably accelerating patient  flow, improving quality, and saving caregiver time, is having the ability to spot situations and risks early, so caregivers and expediters can intervene in the moment.  Real-time, contextual information that is simple to digest and easy to access, is the currency with which to make this happen.  By combining clinical expertise, real-time and predictive analytics, and pre-defined action sets, Humber River Hospital in Toronto, Canada, is unlocking capacity, improving protocol compliance and reducing patients sent to ICU, while at the same time improving care team communication and reducing caregiver stress.

OSF Healthcare has been building and deploying predictive models into operational workflows for more than 10 years.  This talk will provide a brief overview of the various methods successfully used to move models into production.

Effective capacity management for surgical departments includes models for prediction of surgery duration, ICU-bed type and length-of-stay. We used deidentified data from 3,000 cardio-thoracic surgeries to develop and validate predictive models of surgery duration that are based on random forest, extreme gradient boosting and linear regression.  The ensembled predictive model of acute cardio-thoracic surgery duration reduced the number of surgeries “behind-the-schedule” by 28% (from 60% to 32%) and boosted the surgery “on-time” by 15% (from 30% to 45%). Surgery planners could benefit from the predictive models by creating an optimized surgery schedule as a prerequisite to effective capacity management and improved patient and staff experience.

We explored the feasibility of deep learning algorithms to improve the accuracy of predicting daily emergency hospital visits by tracking their spatiotemporal association with PM concentrations. We compared predictive accuracy of the models based on PM datasets from a single but more accurate air monitoring station in each district and multiple but less accurate monitoring sites within a district in Seoul, South Korea. We used MLP (multilayer perceptron) to integrate PM data from multiple locations and then LSTM (long short-term memory) models to incorporate the intrinsic temporal PM trends into the learning process. The results reveal evidence that predictive accuracy is improved from 1.67 to 0.79 in RMSE when spatial variations of air pollutants from multi-point stations are incorporated in the algorithm as a 9-day time window. The findings suggest guidelines on how environmental and health policymakers can arrange limited resources for emergency care and design ambient air monitoring and prevention strategies.

The ultrasound examination is one of the most common techniques of medical imaging. FOLLISCAN is an analytical and predictive system based on interpretable deep learning algorithms to support healthcare practitioners in ultrasound ovaries diagnostics - antral follicles examination. The counting of follicles is of major importance, e.g., it helps estimate the ovarian reserve. Moreover, a large number of antral follicles indicates polycystic ovarian morphology.

FOLLISCAN is designed for fertility clinics, diagnostic centers and, hospitals and will respond to their well-identified need, i.e. time and cost savings - through wider access to ultrasound diagnostics to make better and more informed clinical decisions. The results of the project address the need to perform ultrasound examinations in a faster, easier, more objective and accurate manner. 
In this talk, we will describe the process of building this solution and show the most important issues for creating a solution based on deep learning.
FOLLISCAN is currently tested and implemented in all INVICTA fertility clinics across Poland.