Machine learning for more accurate predictions and more precise targeting

Predicting consumer behavior can be a complex process, especially in a healthcare setting. Fortunately, the abundance of data sources (such as claims data and clinical statistics), provide ample opportunity to generate meaningful insights. The advancement of machine learning algorithms has opened up even more opportunities to get ahead of complex problems and to predict future behavior.

At NextHealth, we have incorporated machine learning into our platform. For example, the platform can accurately predict emergency room visits by identifying health risks and utilization patterns within identified member populations. Backed by these predictions, the platform then targets at-risk and impactable member clusters, assigns them to the most effective intervention campaigns, and measures what works for whom.

“By integrating machine learning into our platform, we can transform health data into actionable intel and impactful results.”

Better Targeting of Populations

Once a client’s raw data has been standardized and processed, NextHealth builds the algorithms to maximize predictive power. Our data scientists mainly use tree-based learning algorithms due to their regression and classification capabilities, along with their scalability into both linear and nonlinear relationships.

In order to target high risk populations that offer the greatest opportunity for impacting a particular use case, we use Decision Trees (a flowchart-like structured algorithm) due to the intuitive visualization design and interpretation capabilities. As the decision trees are processed, we split the features based on the largest information gain. The output gives us the difference between the impurity of the parent nodes and the sum of the impurity of the child nodes. This splitting process is repeated at each child node until it reaches to the very bottom leaves. In order to avoid overfitting – a very deep tree with lots of nodes – we use a “prune” technique to limit the maximum depth of the tree.

An example from the platform: In this case, decision trees help us predict the population clusters that have the highest ER utilization in comparison to the total population and offer the highest potential for use case impact

Member-Level Risk Prediction

At the member level, we run our risk prediction by using Random Forest and Gradient Boosting.

Random Forest is an ensemble of decision trees and applies a Bagging technique to tree learners. Random Forest takes random samples of the training set (with replacement) and grows a decision tree from each sample. After many trees are formed, the prediction will be aggregated and decided on majority vote. The ensemble method is believed to be robust to noise.

Gradient Boosting is an additive model for including weak learners using a gradient descent procedure – an iterative method that takes steps proportional to the negative of the gradient of the function to find a local minimum of a function. Decision trees are used as the weak learner in Gradient Boosting. Trees are implemented one at a time, and a gradient descent procedure is used to minimize the loss when adding trees.

While Random Forest reduces variance, Gradient Boosting reduces bias. We use a stacking technique to combine the two algorithms for improved accuracy.

The Result: Better Predictions, Improved Targeting

At NextHealth, we deliver measurable outcomes for our health plan customers by 1) better targeting their most impactable members, 2) deploying the personalized and persistent interventions most likely to impact behavior, and 3) measuring and optimizing what works for whom. The machine learning methods we employ drive more reliable behavioral predictions and more accurate targeting of at-risk member populations. Ultimately, these processes enable health plans to get ahead of costly behaviors and deliver the most impactful interventions to the members who are most likely to be receptive. By integrating machine learning into our platform, our customers transform big data into actionable intel and impactful results.

By Cathy Zdravevski, Data Scientist, NextHealth Technologies

Machine learning helps us get ahead of many problems. Explore other use cases.