Background: Ankylosing spondylitis (AS) is an inflammatory arthritis that affects the spine, frequently presenting as chronic back pain.The Bath Ankylosing Spondylitis Disease Activity Index (BASDAI) is a validated, patient-reported measure for assessing AS disease activity. The BASDAI is widely used in clinical trials and is appropriate for use in routine clinical practice. Consistent capture of disease activity over time is important for monitoring changes in AS disease activity and need for treatment. Yet, documentation of the BASDAI is inconsistent in real-world data (RWD) sources such as electronic medical records (EMRs). This limits the potential role of these data sources for supporting large, heterogeneous research studies on AS treatment and outcomes.
Objectives: This study aimed to validate a machine learning model to generate estimated BASDAI score categories at specific timepoints for AS patients using clinical notes from a RWD source.
Methods: The estimation model was an ordinal regression based multivariable model with features generated from the clinical notes from rheumatologist visits. Data from the OM1 PremiOM Axial Spondyloarthritis (AS) dataset were used to create a training cohort (2,800 encounters from 750 patients) and validation cohort (1,199 encounters from 322 patients), all with recorded BASDAI scores. Spearman R and Pearson R values were calculated to evaluate estimated BASDAI scores versus recorded BASDAI scores on a continuous scale, and the area under the receiver-operating-characteristic curve (AUC), positive predictive value (PPV), and negative predictive value (NPV) were calculated to assess performance of the model as a binary predictor. A cutoff of 4 was used for the binary variable based on the relevance of this score as a measure of active disease. The model was applied to encounters without recorded BASDAI scores in the AS dataset.
Results: The machine learning model estimates BASDAI scores with very good performance. The model had a Spearman R value of 0.63 and Pearson R value of 0.62. The model had an AUC of 0.82, PPV of 0.82, and NPV of 0.72 when evaluating performance using the dichotomized version of the outcome. Application of the model to the AS dataset resulted in the generation of estimated BASDAI score categories for an additional 123,193 encounters from 17,744 AS patients.
Conclusions: At the individual patient level, use of the BASDAI estimation model could provide a more complete view of a patient’s disease activity and response to treatment over time. At the population level, application of the model to RWD sources increases the number of patients and encounters available for research on AS treatment and outcomes.
