Use ACE Scores in Machine Learning to Predict Disease Earlier and Improve Outcomes

Machine learning in healthcare uses data, an algorithm, and a model to predict an event and suggest interventions that can improve the outcome of that event. A machine learning model for a health system could be designed to predict, for example, who in the hospital is likely to get a central line-associated bloodstream infection (CLABSI), which will then allow clinicians to pay special attention to infection control best practices for those identified patients.

Initially, the machine is fed historical data (e.g., the health system’s patient attributes for those that did and did not get a CLABSI over the past two years), which an algorithm uses to learn relationships (e.g., historical CLABSI rates relative to patient age and comorbidities, duration of catheter insertion, and catheter type used).

This is the essence of the machine learning workflow (Figure 1), which stores the relationships and applies them to make the prediction (e.g., given this patient’s similarities to all the historical CLABSI cases, he has a 75 percent chance of getting an infection today). The model is trained with new data as it becomes available (e.g., CLABSI cases over the next six months), which improves the reliability of future predictions.

Figure 1: The components and processes of a machine learning model to predict a healthcare outcome

Health systems can use machine learning to predict sepsis, the likelihood of readmission or missing an appointment, and dozens of other clinical and operational conditions. From the workflow described above, it’s evident that “nutrient-rich” data sources are necessary to feed predictive algorithms in a machine learning model that’s designed to improve health outcomes.

An important health system objective is making accurate predictions, which relies on capturing a data snapshot of the entire patient. Adverse Childhood Experience (ACE) scores fill one of the significant gaps health systems typically don’t have data for.

The Data Headwaters of ACE Scores

From 1995 to 1997, the CDC and Kaiser Permanente conducted a landmark survey of more than 17,000 people to learn about their early childhood experiences (e.g., abuse, relationships, drug and alcohol use, etc.) and current health statuses. Participants were asked about ten types of childhood trauma related to abuse, household challenges, and neglect, and were assigned an ACE score on a scale from 0 to 10. The participants were also asked about personal and family health. The study showed a strong correlation between high ACEs and negative health outcomes later in life, including risky health behaviors, chronic health conditions, reduced lifetime income, and early death.

While national efforts aim to prevent child abuse altogether, much can be done later in life to prevent further consequences from those early experiences. The opportunities for collecting ACE data are relatively rare, either during childhood or a primary care visit later in life if the doctor is involved with a data-collecting program. But ACE scores only need to be collected once; they never change, and, as the CDC puts it, provide tremendous insight into a “person’s future violence victimization and perpetration, and lifelong health and opportunity.”

Using ACE Data to Improve Individual and Community Health

ACE data is typically used in public health programs for state- and community-wide prevention efforts. In the context of machine learning, health systems should use it to benefit individual patients, so they can flag them as high risk, treat them appropriately, and hopefully prevent ACE-related conditions from surfacing. Family Health History and Health Appraisal questionnaires are readily available instruments for establishing ACE scores. These are new sources of information to consider and systems should broadly adopt surveys as instruments for improving population and individual health. The better picture organizations can paint of a person’s health history, the better they can predict the need for future interventions. The better the incoming data, the better the predictions.

Here’s an example of how health systems can use ACE scores and machine learning to improve patient outcomes: a patient named Alex is admitted to the ED at a hospital where a machine learning model is used to predict opioid addiction risk. The model discovers the strong relation between high ACE scores and opioid abuse in the historical data, and it flags Alex as being high risk for addiction. The model reports that Alex’s high ACE score has a strong impact on the prediction, and clinicians act on this information by avoiding an opioid prescription. Clinicians also treat the underlying factors that make Alex more prone to abusing opioids.

Health Systems Need Feature Engineering on All Source Data

It’s possible to have bad data or too much data, so feature engineering separates the wheat from the chaff. Feature engineering encodes data into formats that are useful for machine learning.

For example, if a health system thinks a patient’s addiction risk may have a seasonal component, then it must convert the date column from “August 14, 2014” to “August.” Once the system gets the data, it can select how it will feed it into the model, and whether to toss, keep, separate, or combine certain variables. This applies to all source data, including ACE scores.

Improve Outcomes with ACE Data

As health systems rely more heavily on machine learning, they must keep in mind that a machine learning model is only as good as the data it receives. Machine learning in healthcare depends on high-quality data to improve outcomes; ACE scores help build better data sets. Incorporating “nutrient-rich” data sources, such as information about ACEs into machine learning models, can improve their ability to predict negative health outcomes, therefore allowing for earlier interventions.

Clinical data collection is moving in the right direction. Today, organizations have data that can be used for machine learning on current problems; in the future, ACE scores, eating habits, and lifestyle data will all be combined to predict diseases earlier and significantly improve population health management.

Additional Reading

Would you like to learn more about this topic? Here are some articles we suggest:

  1. How Machine Learning in Healthcare Saves Lives
  2. How Healthcare Text Analytics and Machine Learning Work Together to Improve Patient Outcomes
  3. How Healthcare AI Makes Machine Learning Accessible to Everyone in Healthcare
  4. There Is A 90% Probability That Your Son IS Pregnant: Predicting the Future of Predictive Analytics in Healthcare
  5. Three Approaches to Predictive Analytics in Healthcare

PowerPoint Slides

Would you like to use or share these concepts?  Download this presentation highlighting the key main points.

Click Here to Download the Slides 

Loading next article...