Target association rule mining to explore novel paediatric illness patterns in emergency settings

Abstract

Background and aims To assess the hospitalized sick children admitted to the pediatric emergency department (ED) and to find new patterns of clinical and laboratory attributes using association rule mining (ARM). Methods In this observational study, 158 children with median (IQR) age 11 months and a PRISM III score of 5 (2–9) were enrolled. Hotspot data mining method was applied to assess clinical attributes, lab investigations and pre-defined outcome parameters of children and their association in sick hospitalized children aged 1 month to 12 years. Results We obtained 30 rules with value for outcome as discharge is given attributes as follows: duration of hospitalization > 4 days, lactate > 1.2 mmol/L, platelet = 3.67/μL, dur_ventil = 0 h, serum K = 5.2 mmol/L, SBP = 120 mmHg, pCO2 = 41.9 mmHg, PaO2 = 163 mmHg, age = 92 months, heart rate > 114–159 per minute, temperature > 98 °F, GCS (Glasgow Coma Scale) > 7–14, gas K = 4.14 mmol/L, gas Na = 138.1 mmol/L, BUN (Blood Urea Nitrogen) = 18.69 mg/dL, Diagnosis > 1–718, Creatinine = 1.2 mg/dL, serum Na = 148 mmol/L, shock = 2, Glucose = 144 mg/dL, Mg(i) > 0.23 meq/L, BUN > 6.54 mg/dL. Conclusion ARM is an effective data analysis technique to find meaningful patterns using clinical features with actual numbers in pediatric critical illness. It can prove to be important while analysing the association of clinical attributes with disease pattern, its features, and therapeutic or intervention success patterns.