Machine learning mines EHRs to predict heart failure

The widespread implementation of electronic health records (EHRs) has proved to be a bumpy ride for many. But the sheer amount of data available in digital form carries with it plenty of potential.

Recent work by scientists from IBM and Sutter Health developed artificial intelligence that can uncover pre-diagnostic heart failure through EHRs.

A study, published in Circulation: Cardiovascular Quality and Outcomes, included a model that used 1,684 heart failure cases along with 13,525 sex, age-category and clinic matched controls for modeling purposes.

Experiments were completed to measure how the machine learning predictive modeling performed in relation to five variables:

• Data Diversity: Eight different data types were individually assessed on model performance.
• Prediction window length: varied from 30 days to five years.
• Observation window length: varied from 60 days to five years.
• Data quantity: varied in the amount of randomly selected training data that ranged from 114 to 11,400 patients.            
• Data density: different amounts of data were selected based on the number of encounters in the observation window.

“Model performance was most strongly influenced by the diversity of data, basic feature construction and the length of the observation window,” wrote Kenny Ng, research staff member in the Center for Computational Health and first author of the study. “In raw form, EHR data are highly diverse, represented by thousands of variants for disease coding, medication orders, laboratory measures, and other data types. It seems obvious that some level of feature construction that relies on well-established ontologies should improve model performance.”

The model performed best when window length was below two years, the training data set at least 4,000 patients, data were diverse as possible and data were confined to patients with more than 10 meetings with physicians in two years.

The study showed that guidelines for the amount and type of EHR data are needed to effectively train machine learning programs to predict disease onset.

“There are many possible directions for future work. First, the approach and methods need to be validated on larger patient data sets from multiple healthcare systems and additional disease targets to better understand the generalizability of the data characteristic impacts on predictive modeling performance,” wrote Ng et al.