Researchers from Children’s Hospital Boston sifted through an anonymized database of 560,000 EMR patient records to develop a mathematical model that shows a decent ability to detect cases of domestic abuse.
All patients had visits recorded over at least a four-year period; overall, 1 to 3 percent had an abuse diagnosis on record, depending on the case definition used.
Using data from two-thirds of the patients, a computer model was trained to differentiate those who ultimately received a diagnosis of abuse from those who didn’t, based solely on their history of visits. The variables associated with abuse (such as a higher number of annual visits, mental health diagnoses, and visits for injury) were used to create a predictive model, which was tested on the remaining third of the patients.
Reis [Ben Reis, PhD, of the Children’s Hospital Informatics Program (CHIP) and the Division of Emergency Medicine at Children’s] and his CHIP colleagues Kenneth Mandl, MD, MPH and Isaac Kohane, MD PhD, found that the model was able to identify these patients an average of 10 to 30 months before the diagnosis was made, with high sensitivity and specificity.
The analysis also found some gender differences in the statistical signs associated with abuse. Alcoholism, poisoning and injuries from external causes were more predictive of abuse in women than in men, while psychoses, affective disorders, and other mental conditions were more predictive in men than in women.
Reis notes that the database used may underestimate the true incidence of abuse, which is sometimes undiagnosed or simply not recorded in medical records. “The data set we used to train and evaluate the predictive model has all the challenges of real-world data,” says Reis. “The important finding here is that in spite of these real-world limitations, the model was able to produce useful and reliable results.”
Reis also developed a color-coded visual display to help physicians quickly process large amounts of information, integrating the patient’s entire diagnostic history into one easy-to-digest view (see accompanying images) and displaying an alert if the history is suggestive of abuse. “Our goal is to communicate an overview of a patient’s longitudinal history to the doctor in 10 seconds,” says Reis. “We’re hoping doctors with access to this kind of view will be able to provide better, more-informed care, including detecting certain medical conditions earlier.”
Image: These sample visualizations present a quick, broad overview of two anonymized patients’ diagnostic histories during the four years preceding their abuse diagnosis. Each small colored bar represents a diagnosis recorded at a particular visit (most recent visits at the bottom), grouped into 12 categories (injury, psychiatric, etc.). Each bar’s color denotes the degree of abuse risk it reflects (green = low; yellow = medium; red = high). The blue arrows at right indicate when the model would have detected a high abuse risk, assuming a false-positive rate of 5%. For patient A, with few recorded visits, abuse risk would have been detected 27 months before the recorded abuse diagnosis. For patient B, with many visits, abuse risk would have been detected even earlier–34 months before it was actually diagnosed.
Press release: “Intelligent” Medical Histories Could Flag Domestic Abuse Sooner
Full article in BMJ: Longitudinal histories as predictors of future diagnoses of domestic abuse: modelling study
