galvanize-dsi-capstone
DocReach: Predicting physician specialty from text data
Sushil K Sharma
Overview
A large social media marketing firm wants to target doctors / physicians based on their practice area. For example, A marketing campaign to target Cardiologists for heart related news feed.
Scope
The marketing firm wants to predict physicians under medical specialties based upon procedures performed over the past year
- Phase 1: Develop a binary classifier to predict Cardiologists
- Phase 2: Develop a multiclass classifier to predict top 5 Specialties
Motivation
Dataset
Data has been provided in the form of csv files:
procedures.csv - contains a list of procedures doctors performed over the past year. The columns of this dataset are as follows: physician_id unique physician identifier, joins to id in physicians.csv procedure_code unique code representing a procedure procedure description of the procedure performed (Text Column) number_of_patients the number of patients the doctor performed that procedure on over the past year
physicians.csv - contains a list of doctors and their unique specialty. Specialty is listed as "Unknown" for the doctors need to be classified. physician_id unique physician identifier, joins to id in physicians.csv Specialty: String e.g. Cardiologist
Process
- Data Acquisition
- EDA
- Data Prep
- Model Selection
- Results
Results
Best model: RandomForestClassifier Best threshold: 0.04 Resulting profit: $45.94 Proportion positives: 0.80 precision recall f1-score support False 0.74 0.85 0.79 1272 True 0.82 0.68 0.75 1228 avg / total 0.78 0.77 0.77 2500
Accuracy Score Train: 0.789 Accuracy Score Test: 0.7708
Next Steps
Exolore new features viz. number of patients