OBJECTIVES: Patient diagnosis is very relevant in studies to evaluate patient flows and treatment patterns by indication. However, most data assets lack diagnosis information, which restricts insights into real-world drug utilization. Heuristic expert models based on medical guidelines were developed to predict patient diagnosis, but low accuracy in respiratory indications limits broader use. The objective of this research is to predict patient diagnosis in patient level data assets based on patient characteristics, historical treatment patterns, prescribed drug, therapy duration and prescriber information.   METHODS: We used EMR (electronic medical record) data including indication, prescribed drug, physician and patient specific data dimensions. The target indications of Asthma and Chronic Obstructive Pulmonary Disease (COPD) were selected using ICD10 codes and patients were classified accordingly. Data cleaning was undertaken to exclude patient records with unclassifiable indications. CHAID models (Chi-squared Automatic Interaction Detector) and Machine Learning (Random Forest) were used to predict patient indication using patient characteristics. In the next step, we used the model to predict indication at patient consultation level. Finally, we compared drug utilization patterns for different indications. RESULTS: Data for ~285k patients observed over Q2/2014-Q1/2015 consisting of Asthma (49%), COPD (22%) and Others (29%). Models were estimated on training sample (30%) and validated on holdout sample (70%). The model produced very high accuracy rates, with Machine Learning outperforming CHAID models. In case of Machine Learning, we obtained hit rates for Asthma of 93% in training sample and 85% in holdout sample, while COPD hit rates were 90% in training and 80% in holdout sample. CONCLUSIONS: We developed and validated an algorithm that models diagnosis versus patient characteristic for Asthma and COPD, and allows for indication prediction. Our approach combines the strengths of medical expertise with robust analysis using CHAID and Machine Learning to create a new standard for prediction of patients’ indication.