Development and validation of machine learning models to predict risk of undiagnosed dementia using healthcare claims and electronic health record data
Deborah E. Barnes, Cynthia Benjamin, W. John BoscardinBackground
Approximately half of people living with Alzheimer's disease and related dementias are undiagnosed.
Objective
To develop and validate algorithms that predict risk of undiagnosed dementia using electronic health record (EHR) and/or healthcare claims data.
Methods
Study participants were adult patients aged 65 years or older without evidence of dementia (diagnosis/medication) at baseline in two U.S. data sources: 1) Medicare claims (2010 to 2021); 2) EHR and claims from a primary care network (2016 to 2023). We applied coefficients from an existing, validated EHR-based algorithm to predictors defined using Medicare claims and used machine learning to develop new EHR- and claims-based predictive models. We assessed model discrimination using c-statistics.
Results
Study participants included 8,374,400 Medicare beneficiaries (mean [SD] age, 76 [7] years; 57% female) and 29,983 primary care patients (age: 75 [6] years; 56% female). Model discrimination was good when applying EHR-based coefficients to Medicare claims-based predictors (c-statistic [95% confidence interval]: 0.770 [0.767, 0.773]) and was improved by refitting the model (0.795 [0.792, 0.798]) with a small added benefit from incorporating new claims-based predictors (0.801 [0.799; 0.804]). Similarly, when both EHR and claims data were available, discrimination was improved by refitting the model with a small additional benefit from including new predictors, regardless of the data source (EHR, claims, or either).
Conclusions
A validated EHR-based algorithm predicted risk of undiagnosed dementia in Medicare claims with good discrimination. Model accuracy was improved by refitting and, to a lesser extent, by including novel predictors.