Risk scores are used widely in healthcare. They help providers identify patients at higher risk for certain events and conditions, enabling focused interventions to reduce that risk. These risk scores can also act as great educational tools for providers, enabling them to show their patients the effect of certain lifestyle changes on their health. For example, a physician can show a patient his risk of a heart attack in the next ten years, and then illustrate how quitting smoking can reduce the risk by nearly 50 percent.
These risk scores are not without their flaws. Many existing risk scores can be improved by using your health system’s data and applying machine learning to predict risk. Additionally, most scores also require a clinician to review a patient’s chart for inputs that are found in clinician notes and textual reports – information that cannot be analyzed without a text analytics solution.
We believe risk scores are a powerful tool for data-driven clinical care and are working to apply customized machine learning healthcare algorithms and text analytics to enhance existing risk scores and begin to develop new ones.
Increased accuracy is enabled by the fact that the machine learning algorithms in healthcare examine larger datasets and cast a wider net across potential contributing variables to identify complex patterns of relationships between your patients’ attributes and the outcomes of interest, according to your population’s full dataset.
Contrast this with the typical approach, in which risk scores are based on decades-old studies and cohorts that often have significantly different demographics from your patients. This improvement in accuracy isn’t simply a hypothesis. The industry has repeatedly seen machine learning algorithms beat internationally recognized risk scores:
CHADS₂ is a well-known risk score that predicts the one-year risk of stroke for patients with atrial fibrillation. CHADS₂ was developed based on a 2001 study using a data set with a couple thousand patients that is now approximately two decades old. Even the original study’s principal author, Brian Gage, suspected in 2004 that CHADS₂ could be improved upon. In that 2004 follow-up study, Gage suggested that left ventricular systolic dysfunction (LVSD), cigarette smoking, and moderate alcohol consumption may additionally impact the likelihood of a stroke.
By developing and training a machine learning model, health systems can test whether these features will have an impact on predicting stroke in their populations, allowing them to improve on the original risk scores. CHADS₂ has one more limitation—it is not widely run on patients with atrial fibrillation, as it requires manual calculation by a clinician after reviewing the patient’s history and their notes.
CHADS₂ looks at a number of inputs, including history of stroke or transient ischemic attack (TIA) symptoms to determine the one-year risk of stroke in patients with atrial fibrillation. History of stroke and history of TIA, in particular, are not well coded in the EMR at most health systems, unless the acute event was treated by the hospital. However, this history is commonly found in physician notes.
Using text analytics, systems can efficiently examine the contents of such notes. They can not only identify notes in which “TIA” is mentioned, but also distinguish between a reference to a patient with TIA symptoms and a patient named “Tia”—or a mention that TIA was ruled out by a clinician. Through a careful application of text analytics, we can identify the patients who truly have a history of TIA.
The same text analytics may be leveraged to help validate whether Gage’s 2004 hypotheses on LVSD, cigarette smoking and moderate alcohol consumption are correct. All three elements are not often entered into discrete fields within health systems’ EMRs, especially the indicator of LVSD. By pulling these values out of clinical notes and diagnostic testing reports, they can be fed into a machine learning model. That model will determine whether those three features, previously locked in free text, have an impact on whether a patient is at risk for stroke.
Testing the hypothesis that incorporating additional risk factors, such as left ventricular systolic function, moderate alcohol consumption, and cigarette smoking could achieve an even better predictive accuracy beyond CHADS₂ becomes feasible in real time by leveraging machine learning and text analytics together.
Text analytics can automate the discovery of data elements essential to the new model, which are normally obscured in text and require manual extraction by the physician if you want to inform clinical decisions, or retrospectively by chart abstractors. Amount of alcohol consumed, ejection fraction, cigarette smoking, and history of TIA are all features often missing from structured coded elements of the medical record and captured most reliably in text reports or provider notes. Machine learning can incorporate and evaluate the association of all of these features with incidence of stroke across a population and analyze the strength and relative contribution of each proposed risk factor, new or existing, to the overall risk, alone or in particular combinations.
By combining machine learning and text analytics, we can improve the accuracy of risk scores, and do so in an automated fashion to identify every patient’s level of risk in time to inform physician decisions about each patient’s care; not simply those for whom the provider has time to collect information and manually run a risk score, and not simply in retrospect through manual chart abstraction.
In an industry that’s historically been limited to one-size-fits-all risk scores that could only run on discrete, structured data, now health systems can run machine learning models, with input from text analytics wherever necessary, to provide tailored risk predictions on both structured and unstructured data.