How Machine Learning in Healthcare Saves Lives
Machine learning and its predictive power enhances our everyday lives: Netflix and Amazon recommendations, spam and fraud detection, mobile check deposits, driving directions, optimized search results, etc. The list grows longer by the hour.
What is machine learning and how does it work? It’s a field of computer science that uses pattern recognition to identify historical relationships in large data sets using an algorithm to create a generalized model of behavior. This model can be used to predict future events—a predictive model.
The “machine” is the computer and software; the “learning” comes from the fact that an algorithm embedded in the software learns from historical data (the values associated with specific input features/variables) and then applies that learning on new data in the form of predicting an outcome—and then cycles through that process iteratively. Via machine learning and a health system’s historical data, the model and predications can be tailored to its business questions.
Given machine learning’s ability to make sense of vast quantities of data, it’s no wonder health systems are desperate to understand how it can help their clinical teams and improve their patients’ outcomes. If machine learning can benefit the average American in myriad ways every day, then can it benefit healthcare?
This article explains why the answer is “yes”—and why machine learning in healthcare will be one of the most important, life-saving technologies ever introduced.
An Industrywide Shift: From One-Size-Fits-All to Customized Machine Learning Models
In the past, simple models based on limited sets of data ruled the day. Simple scoring systems were devised to make the models easy to implement. A popular example is the LACE index, a simple calculation that provides the likelihood of patient 30-day readmission risk. It’s based on length of stay (LOS), acuity of the admission, patient comorbidities, and ED visits within the last six months. In each of these categories, points are assigned—a LOS of three days equals three points, for example. Points from each category are added up to form the LACE index.
The LACE index is indicative of how healthcare has worked for the last 20-30 years: a national study generates guidelines and a simple calculation to help systems prioritize which patients are most at risk of something. This outdated approach is problematic for two key reasons:
- Models based on simple rules and data sets that don’t generalize well don’t yield accurate predictions for organizations that implement them. LACE was developed from patients seen in Ontario from 2004 to 2008. Chances are, your patient demographics don’t closely match those in Ontario. Using a model based on populations whose demographics may be very different can lead to poor predictive value because the models are trained to look for factors in the data.
- Reactive vs. proactive. LACE requires the patient’s LOS, so the score is only available at discharge. Systems need a more proactive approach, with risk scores calculated earlier in a patient’s stay.
Machine learning fills the gaps left by the simple models of the past. It learns the important, and sometimes unique, relationships in health system data on its patients and their outcomes under present day conditions, so the system doesn’t have to rely on scores made on other populations 20 years ago. Machine learning allows health systems to break out of literature-based models, deliberately limited to relatively simple constructs and finite numbers of potential contributing factors to allow clinical teams to find the data elements and perform the calculations manually. Instead, machine learning allows for algorithms to consider numerous potential contributing factors, and rapidly generate and test more complex predictive algorithms.
Another past-to-present differentiator: the volume of data produced by healthcare systems (e.g., patient notes, lab data, time stamps, medication information, radiology images, scheduling data, and billing data) is continually growing. Clinicians have so much data to read, interpret, and analyze.
Machine learning processes this data to provide summaries and metrics that help healthcare providers perform their work more effectively and accurately. The opportunities for machine learning to improve clinical, workflow, and financial outcomes are limitless.
Limitless Opportunities for Machine Learning in Healthcare
From reducing readmissions to predicting a patient’s propensity to pay, machine learning does more than just benefit healthcare—it is a life-saving technology health systems and their patients desperately need. Although the improvement opportunities are limitless, there are several high-profile clinical and financial challenges machine learning can help healthcare organizations overcome:
Many health systems are penalized if their 30-day readmission rates are too high. Machine learning can reduce readmissions in a targeted, efficient, and patient-centered manner. Clinicians receive daily guidance as to which patients are most likely to be readmitted and how they can reduce that risk.
Preventing hospital acquired infections (HAIs)
Health systems can reduce HAIs, such as central-line associated bloodstream infections (CLABSIs), by predicting which patients with a central line will develop a CLABSI. Clinicians can monitor high risk patients and intervene to reduce that risk by focusing on patient-specific risk factors.
Reducing hospital length of stay (LOS)
Health systems can reduce LOS and improve other outcomes, such as patient satisfaction and red census alerts, by identifying patients that are likely to have an increased LOS and then ensure that best practices are followed to reduce prolonged LOS while improving outcomes.
Predicting chronic disease
Machine learning can help health systems identify patients with undiagnosed or misdiagnosed chronic disease, predict the likelihood that patients will develop chronic disease, and present patient-specific prevention interventions.
Reducing one-year mortality
Health systems can reduce one-year mortality rates by predicting the likelihood of death within one year of discharge and then matching patients with appropriate interventions, care providers, and support.
Uncompensated care is a growing problem for most health systems. With machine learning, health systems can determine who needs reminders, who needs financial assistance, and how the likelihood of payment changes over time (and after events).
It’s hard to maintain schedules that keep clinicians and patients happy. Slots are often over- or under-booked because someone showed up late, didn’t show up, or showed up without warning. With machine learning, health systems can create accurate predictive models to assess, with each scheduled appointment, the risk of a no-show, ultimately improving patient care and the efficient use of resources.
Improving sepsis outcomes
Mortality rate from sepsis is high: anywhere from 10 percent to 52 percent, depending on the study. Fortunately, there are several ways machine learning can be used to improve sepsis outcomes:
- Early, in-house recognition. Predictive models can identify patients at a higher risk of developing sepsis while in the hospital, allowing clinicians to start treatment earlier.
- Predicting complications, deterioration, or need for escalation of care. Predictive analytics can calculate the likelihood that patient would need ICU treatment, intubation, or extended LOS, and help stratify patients by risk.
- Predicting readmission due to sepsis. Having a predictive model that identifies patients at higher risk for readmission enables clinicians to provide early follow up and close contact, decreasing the likelihood of patients being readmitted.
Considering the resource constraints within most health systems, machine learning is crucial to efficiently deploying resources toward achieving business goals (e.g., reducing readmissions, reducing one-year mortality, and preventing HAIs).
A Deeper Dive into Machine Learning: The Importance of Historical Data
At its most basic definition, machine learning refers to a group of algorithms that learn from data. These algorithms work using patterns extracted from historical data, rather than fixed formulas. Basic machine learning can help health systems answer a broad array of questions, as long as there’s good historical data.
For example, if you went to the hospital for flu-like symptoms, then a clinician thinking along the lines of traditional algorithms might say, “You have a fever, aches, general weakness, and no cold symptoms. This looks like the flu.” A clinician thinking along the lines of machine learning might say, “Your symptoms are the same as 50 recent patients who had the flu. You probably do too.” Using machine learning, clinicians use historical data to help make an accurate diagnosis.
This scenario is an example of a classification machine learning problem. A classification algorithm gives a probability score of a person having the disease or the probability of an event happening:
- What is the likelihood that a chronic obstructive pulmonary disease (COPD) patient will be readmitted within 90 days of discharge?
- What is the likelihood that a person will no-show for their appointment?
These questions are posed with a lot of example data and the expectation that a model will give a probability from zero (low) to one (high). The predictions are made based on matching the pattern of each patient’s characteristics (e.g., potential contributing factors) to the pattern of patients who historically were vs. were not readmitted within 90 days of discharge.
A regression algorithm, on the other hand, predicts a continuous value (e.g., how many days a patient needs to stay in the hospital or how much money a patient will cost the system over the next year).
How One Large Health System Improved Care for COPD Patients with Machine Learning
After identifying patients with COPD as a top readmission diagnosis, one large health system turned to machine learning for help. COPD is a common chronic disease (responsible for approximately 135,000 deaths annually), punctuated by preventable exacerbations often requiring admission. The national cost for one COPD readmission is $9-12k and 20 percent of COPD patients are readmitted within 30 days.
Despite implementing evidence-based order sets, interdisciplinary plans of care, and patient navigators to assist with transitions of care, the system lacked the ability to ensure it was targeting higher risk patients (and couldn’t measure the work to identify which interventions were benefiting the patients). It knew it needed to get upstream of the readmission event, accurately predicting which patients were at greatest risk for readmission and providing the opportunity to intervene, so it partnered with Health Catalyst® to build a predictive model for risk of readmission for patients with COPD.
Now the health system can identify which patients are at greater risk for readmission and understand the factors increasing the risk. To mitigate this risk, it can evaluate the effectiveness of the interventions and identify which activities are contributing to a reduced rate of readmission. The system’s pulmonary navigators can easily access the predictive model data and utilize the predictive scores to identify which patients with COPD have a higher risk for being readmitted, allowing them to prioritize their work, assess the patient’s needs, and optimize the transition from hospital to home.
Unlike historical methods for developing risk calculators and treatment algorithms, machine learning allows for an entirely new approach to understanding and visualizing patterns in the data, by generating and retraining ever-evolving models to describe associations found in the data set—which is refreshed every 24 hours or less, then expressing these associations as predictions of risk for an outcome of interest for a particular patient.
Machine Learning can incorporate a greater number of variables into the analysis, offer insight into relative weights of contribution for each risk factor, and perhaps, point out new associations between available patient characteristics and the outcome of interest—all while permitting the clinical team to have visibility into the uncovered patterns so the clinician can take these risk profiles into consideration while planning and prioritizing the care team’s areas of focus to improve each patient’s clinical course.
Machine Learning Is the Life-Saving Technology Healthcare Needs
By delivering accurate, timely risk scores, enabling precise resource allocation, and driving lower costs and improved outcomes, machine learning helps saves lives. As health systems continue to replace burdensome, manual processes with faster, more scalable technology solutions, the widespread use of machine learning becomes an industrywide imperative. Although barriers to achieving this goal exist, health systems can start today by understanding machine learning basics and beginning to explore how it can help their patients, clinical teams, and bottom lines.
Would you like to learn more about this topic? Here are some articles we suggest:
- The Top Three Recommendations for Successfully Deploying Predictive Analytics in Healthcare
- How Healthcare AI Makes Machine Learning Accessible to Everyone in Healthcare
- Three Approaches to Predictive Analytics in Healthcare
- There Is A 90% Probability That Your Son Is Pregnant: Predicting the Future of Predictive Analytics in Healthcare
- DKA Risk Prediction Tool Helps Reduce Hospitalizations