Reducing Readmissions with Machine Learning

Can futuristic-sounding technologies really drive meaningful outcomes that improve the everyday performance of health systems?

You’ve no doubt heard the prediction that machine learning and artificial intelligence will transform healthcare. But can these futuristic-sounding technologies really drive meaningful outcomes that improve the everyday performance of health systems?

The answer to that question may lie in one of the biggest thorns in the side of every health system — the need to reduce preventable 30-day readmissions. Inducements to make a dent in readmissions are powerful but most health systems have no clue where to start. They often implement guidelines and simple calculations based on national studies to prioritize patients with the highest risk of readmission—a common method that doesn’t necessarily offer impressive results, as the rules draw from a small subset of data and can only be applied to a narrow set of patients.

Machine learning (ML) algorithms, however, have the potential to start a new paradigm in healthcare risk modeling by drawing from huge datasets and being applied to every patient.

How Machine Learning Reduces Unplanned and Avoidable Readmissions

ML can reduce unplanned and avoidable readmissions in healthcare by leveraging historical relationships and trends in the data to develop analytic models and algorithms to produce actionable predictions. ML learns the important relationships in data from past patients and outcomes during the last few years to generate a predictive readmission risk model tailored to a specific population within a specific health system. The readmission risk model leverages whatever data is available when the prediction is needed so that limited resources can be focused accordingly.

Health systems are increasingly using ML to better prioritize at-risk patients and optimize care decisions. Health Catalyst developed a new ML tool called—a free, open-source predictive analytics software package that automates and democratizes ML. The goal is to make it easy for health systems to create predictive models that are specific to their data to reduce readmissions, calculate infection risk, and much more. ML can help transition health systems from relying on retrospective analytics to leveraging predictive and prospective analytics.

A Machine Learning Principle to Live By: The Five Rights of Information Delivery

When it comes to implementing ML in healthcare, health systems should operate according to an important guiding principle—the five rights of information delivery. They should strive to provide the right information to the right audience, at the right granularity, at the right time, in the right visualization/modality. This principle facilitates the right actions to improve outcomes, similar to the five rights of medication administration, which are known to reduce medication errors and patient harm.

#1: The right information

In the context of reducing readmissions, ML identifies patients with the highest risk of readmission. It can also isolate the patient-specific modifiable risk factors that lower the patient’s risk, and non-modifiable risk factors that should be considered when determining appropriate interventions. ML can also help differentiate groups of patients of similar risk based on the interventions that they will benefit most from, ultimately guiding clinicians as to which interventions are likely to have the broadest impact on the population. A strong ML model can provide the right information needed for patient-centered decision making in a concise and timely manner.

#2: The right audience

When developing a truly valuable ML model, it’s crucial to know the audience. In part, this is likely the same interdisciplinary health system team that identified the need and use case for predictive analytics to begin with. Health systems should include people directly involved in front-line interventions from the beginning to ensure strong adoption of the solution and trust in the underlying variables, logic, and decisions. Identifying the right audience is as important as creating an ML model that provides the right information. They understand how to pair their deep domain knowledge with the information to intervene with the patients at highest risk, who are most likely to benefit from the chosen interventions.

#3: The right granularity

Healthcare data is recorded and stored in many different sources and levels of granularity. The data can relate to the individual patient, visit, or specific event (e.g., lab test). ML considers how all the different variables interact with one another and contribute to an increased risk of readmission. It can consume data from multiple granularities and deliver insight at the right granularity in the form of a patient-specific predicted probability of being readmitted. Additionally, ML can help uncover more granular modifiable-risk factors that can substitute for a non-modifiable risk factor; patient age provides a good example of this. While a clinician cannot modify a patient’s age to lower their readmission risk, age might be serving as a proxy for more granular age-related risk factors that the clinician can work to address. Examples of age-related risk factors might be the frailty of the patient, an increased fall risk, or the patient’s living arrangement.

#4: The right time

There are multiple opportunities to intervene with a patient to reduce their risk of readmission (e.g., hospital admission, discharge, after discharge, etc.). The right time to capture variables and make predictions depends on the specific use case and associated interventions. For example, identifying high-risk, in-hospital patients as soon after their admission as possible allows ordering providers and hospital staff to intervene and lower a patient’s readmission risk. Identifying high-risk patients at discharge, or shortly after, allows hospital support staff, transitional services, and primary care providers to intervene outside traditional inpatient channels to also lower a patient’s readmission risk. The type and amount of information available for a readmission risk model to leverage is different at each of these varying points in time.

#5: The right visualization and modality

A clinician’s time is valuable, so it is important to incorporate the insight and guidance from the readmission risk model into their regular course of work. By deploying a readmission risk model such that the output is embedded in existing workflow at the point of care, the clinician does not have to spend time searching for information, they can instead evaluate and act on the information at their fingertips. Pairing best practice interventions and potential health outcomes with an alert on a patient’s chart in the EHR or a daily worklist can help validate a clinician’s intuition and result in better and more consistent care. Identifying at-risk patients and delivering actionable information to clinicians and transition services in the right visualization and modality breaks down many of the barriers to adoption.

A New Paradigm—Rules are out, Machine Learning Is in

Traditionally, healthcare risk models have been treated as a type of “locked box” in which only a select few have access to variables and logic contributing to the risk calculation and output. Machine learning has the potential to transform this norm by drawing from the entire EHR and other data sources, at all granularities, and applying insight towards all patients. Leveraging algorithms that draw from a complete set of information is the only way to make informed, forward-looking, patient-centered decisions. ML models can be transparent and trustworthy, and interdisciplinary health system team members are more likely to trust the information and make decisions based on data. An additional benefit to this approach is the opportunity to refine the readmission risk model over time, and replicate, adjust, and repurpose it to fit additional use cases.

ML provides a dynamic layer of intelligence for health systems, and if implemented appropriately, can be used to reduce readmissions in a targeted, efficient, and patient-centered manner.

Levi Thatcher, Ph.D. is Director of Data Science for Health Catalyst (, a leader in healthcare data warehousing, analytics and outcomes improvement.