Random Forests Application In Missing Data And Predictive Modelling For Hierarchical Routine Clinical Data: A Case Study Of Childhood Pneumonia In Kenya

Abstract

Health stakeholders usually need complete, accurate and reliable estimates of various health outcomes to make decisions on improving health care delivery. Missing observations especially in clinical routine data is one of major setbacks in evaluating public health problems efficiently. One of the tools used to measure clinical quality is the Paediatric Admission Quality of Care (PAQC) score. We seek to identify factors that in uence clinical quality in this study after dealing with missing values.The main objective of this study is to identify key determinants of Pediatric Admission Quality of Care (PAQC) score using Random Forests. Data on a total of 2027 children between 2 and 59 months who were admitted in selected county hospitals in Kenya was used. The data contained clinical data from admission to treatment. Random forests missForest package was used to impute missing data. Cumulative logit mixed models were fit with PAQC score as an outcome and age, sex, comorbidity, weight, clinician sex and cadre, hospital workload, malaria prevalence, intervention arm and time of admission as predictors to determine the signi cant determinants of clinical quality. The models were nested within both hospital and clinician levels. Both Random forests and conditional random forests were used to determine variable importance. The cumulative logit mixed model nested within both clinician and hospital level was selected based on AIC. Weight of the child, clinician sex, cadre and the time of admission were significant determinants of PAQC score based on the P values at 0.05 level of signficance. A unit increase in weight increases the probability of a higher PAQC score by 0.06, while being attended by a medical o cer relative to a clinical officer increases the probability by 0.27. The time of admission increases the probability by 0.11. On the other hand, PAQC scores would be lower if the clinician was male. The probability of a reduced PAQC score if a clinician is male is 0.49. Month, weight, intervention arm and hospital workload were the most important variables in predicting the quality of care while age and the number of comorbiditieswere the least important using Random forests models. Based on the cumulative logit mixed models, the study concludes that hospital level, weight of the child, clinician sex, cadre and the time of admission are key determinants of PAQC score. On the other hand, age and the number of comorbidities for a given patient may not strongly influence the quality of care provided based on the random forests models. The mechanisms around these associations however need to be studied extensively. Pneumonia