Agenda

Come join us to learn how one of UnityPoint's affiliates revamped their re-admissions reduction strategy using predictive modeling across the care continuum to reduce risk-adjusted re-admissions by 40% over three years. Under the hood are predictive models that forecast hospital length of stay, 30 day readmission risk, a heat map depicting likelihood of readmission each day post discharge, and follow up visit no show risk at the patient level each morning. We'll share the integrated tool, development details, and how use has spread organically across our health system.

Healthcare reporting and analytics has traditionally treated free text notes as a black hole. The Advanced Analytics team at OSF Healthcare has been leveraging natural language processing (NLP) to address this opportunity space for more than a year. Projects have included the extraction of Ejection Fractions from Echocardiograms, concept analysis of patient satisfaction surveys and named entity recognition for PHI and ailment identification. This session will provide a practical overview of NLP approaches the team has found useful and provide tips on how an organization can get started mining intelligence out of its free text without requiring large scale investments.

Pharmaceutical industry operations are a complex network of interrelated business entities. There is an opportunity for fraud at each stage in the manufacturer/supplier-to-prescriber-to-dispenser-to-consumer distribution chain, and there are many known fraud schemes. We focused on the prescriber-dispenser-consumer end of the chain, and in particular, on schemes involving controlled prescription drugs. 

Creating features from heuristics (domain experts) and deduction (graph analytics), which we employed in machine learning models, we developed a tool that can detect prescriber  fraud schemes such as collusions, negligent prescribing and pill mills.  Taya Fernandes, Drug Diversion & Data Mining Analyst, Medicaid Fraud Control Unit, Indiana, contributed to this presentation.  

People with kidney failure treated by dialysis exhibit higher hospitalization rates than those without kidney disease, resulting in higher morbidity and mortality rates. Fresenius Medical Care developed, tested, and deployed a predictive model (with over 200 clinical and non-clinical variables) to identify patients at risk of hospitalization in the next year. Interdisciplinary clinical teams performed interventions with the identified high-risk patients. To date, we have observed a 23% reduction in average yearly hospital admission rates compared to controls in a regional assessment. This presentation will discuss the creation, implementation, and challenges with large scale deployment of this predictive model.

For children who have mental health conditions, starting therapies early is key for their future quality of life. Early and accurate screening tools are vital.

We present a clinically validated machine learning driven App to screen for autism. We give an overview of the development and optimization of two machine learning algorithms that use different media to identify autism, as well as the combination of their results, and the clinical validation of the algorithm's performance.  We also present an algorithm to simultaneously screen for multiple conditions. We show its preliminary expected performance when outcomes of ADHD or autism are possible.

Successful care management (CM) programs begin with identifying the right people to manage. Identifying the wrong individuals can lead to a costly and inefficient allocation of limited CM resources. At CDPHP, we’ve developed a cloud-based framework for population health management using predictive models to identify members who would benefit from CM based on their prospective need for such services. Additionally, the models aid in triaging members to ensure they receive the resources best suited to their needs. Utilizing potential CM needs increases the opportunity to positively impact members’ health and limits false positives compared to traditional methods.

The ability of software to reason, answer questions and intelligently converse about clinical notes, patient stories or biomedical papers has risen dramatically in the past few years. This talk is intended for practicing data scientists and software engineers, and covers state of the art natural language processing, deep learning, and machine learning libraries in this space. We'll share notebooks & benchmarks from real projects using open-source software packages spaCy, Spark ML, TensorFlow, QuickUMLS, SyntaxNet and Spark NLP. The covered use cases will include clinical data abstraction, patient risk prediction, negation & temporal scope detection, topic modeling, and research paper classification.

MultiCare Health Systems’ OR administrators spent an inordinate amount of time sifting through a mountain of data to find insights about room utilization, number of cases, first case on-time starts, turnover, and other key performance indicators metrics. Building custom dashboards was labor-intensive and often didn’t contain all the information necessary for decision making. MultiCare found that having one source of truth that informs perioperative leaders and can “push” actionable insights is necessary to keep up with an ever-changing healthcare landscape. This session will detail how predictive analytics, machine learning, and cloud and mobile technologies helped MultiCare improve OR Operations.

The Low Carb Program is redefining chronic disease and wellness with over 275,000 global members and clinically-reviewed outcomes demonstrating 40% of people eliminate a diabetes medication and 1 in 4 completers reverse their type 2 diabetes at 1-year. The platform is global and engagement is completely AI-led, personalised to each patient and driven through real-time data feedback and insights. 

Accurate data is necessary for purposes ranging from clinical decisions to service billing to scientific studies. However, with large volumes of data, error detection becomes difficult. To direct the efforts of coders poring through medical records looking for discrepancies in comorbidity reporting, the Data Science team at Fresenius Medical Care developed methods to detect patients likely to have unrecorded comorbidities. By using lab test results, natural language processing of rounding notes, and patient demographic information, we were able to improve the results of coders ten-fold. This session will discuss the development of these methods and the challenges along the way.