Sequential pattern mining identified patterns of healthcare encounters preceding diagnosis of juvenile arthritis in a Canadian province’s health administrative data

Adults with rheumatic diseases have increased healthcare encounters years before diagnosis compared to the general population; whether this is true in juvenile arthritis (JA), the most common childhood-onset chronic rheumatic disorder, is unknown. Our objectives were to (a) identify patterns of healthcare encounters before diagnosis for JA patients, and (b) test for differences in healthcare patterns between JA patients and matched controls without JA. A registry of 386 JA patients from the only pediatric rheumatology centre in Manitoba, Canada (1980-2012) was linked to provincial health administrative data (HAD). Each JA patient was matched to five controls by age, sex and healthcare region. International Classification of Disease codes were mapped to Clinical Classification Software group codes. Encounters were binned into 6-month intervals from birth until diagnosis/index date. PrefixSpan , a sequential pattern mining algorithm, identified time-ordered sequences of healthcare encounters (patterns) shared by ≥ 30 individuals. Results were processed by 1) estimating p-value upper bounds for patterns associated with JA; 2) applying the Benjamini-Hochberg method to limit false discovery; 3) calculating set and sequence differences; and 4) applying k-means clustering to sequences. The ability of patterns to discriminate between JA patients and matched controls was assessed using odds ratio (OR) and 95% confidence intervals (CI). 386 JA patients (diagnosed 1983-2012) were matched to 1930 controls, 65.3% females. Median age of diagnosis for JA patients was 8 (3-12, 25 th -75 th percentile) years. We clustered 255 patterns into 28 groups. Twelve patterns were unique to JA patients. The median number of healthcare encounters in patterns was 3 (2-3, 25 th -75 th percentiles). The most common pattern contained diagnoses for joint disorders up to six months before JA diagnosis; this pattern discriminated JA patients from controls (OR 140.91; 95% CI: 90.17 - 218.90). 80% of patterns included an infection. SPM applied to HAD discovered JA-discriminating patterns of healthcare encounters. These patterns demonstrated face validity, resonating with the current understanding of the biological basis of inflammatory arthritis. In the future, validated patterns could be incorporated into healthcare systems to flag at-risk patients, prompting timely rheumatologist assessments for earlier diagnosis and preventing long-term morbidities.