python-diabetes

Overview

Simple xgboost model python model based on the medical records for Pima Indians and whether or not each patient will have an onset of diabetes within ve years.

A sample notebook is located here.

Datasets

The dataset required to train or evaluate this model is the PIMA Indians Diabetes dataset available here. The teradataml code to import it is

import pandas as pd
from teradataml import copy_to_sql

df = pd.read_csv("http://nrvis.com/data/mldata/pima-indians-diabetes.csv", header=None)
df.columns = ["NumTimesPrg", "PlGlcConc", "BloodP", "SkinThick", "TwoHourSerIns", "BMI", "DiPedFunc", "Age", "HasDiabetes"]

copy_to_sql(df = df, table_name = "PIMA", index=True, index_label="PatientId", if_exists="replace")

CREATE TABLE PIMA_PATIENT_FEATURES AS 
    (SELECT 
        patientid,
        numtimesprg, 
        plglcconc, 
        bloodp, 
        skinthick, 
        twohourserins, 
        bmi, 
        dipedfunc, 
        age 
    FROM PIMA 
    ) WITH DATA;
    
    
CREATE TABLE PIMA_PATIENT_DIAGNOSES AS 
    (SELECT 
        patientid,
        hasdiabetes
    FROM PIMA 
    ) WITH DATA;
    

Training

The training.py produces the following artifacts

• model.joblib (sklearn pipeline with scalers and xgboost model)
• model.pmml (pmml version of the xgboost model and sklearn pipeline)

We also use save a global explainability plots from xgboost which help understand the importance and contribution of each feature.

Evaluation

Evaluation is defined in the evaluate method in scoring.py and it returns the following metrics

• Accuracy
• Recall
• Precision
• f1-score

We produce a number of plots for each evaluation also

• roc curve
• confusion matrix

We also use shap to save some global explainability plots which help understand the importance and contribution of each feature

• shap feature importance

Scoring

This demo mode supports two types of scoring

• Batch
• RESTful
• In-Vantage (PMMLPredict)

In-Vantage scoring is supported via the PMML model we produce during scoring.

Batch Scoring is supported via the score method in scoring.py.

The following table must exist to write (append) the scores into

CREATE TABLE pima_patient_predictions (
    job_id VARCHAR(255), -- comes from airflow on job execution
    PatientId BIGINT,    -- entity key as it is in the source data
    HasDiabetes BIGINT,   -- if model automatically extracts target 
    json_report CLOB(1048544000) CHARACTER SET UNICODE  -- output of 
 )
 PRIMARY INDEX ( job_id, PatientId );

RESTful scoring is supported via the ModelScorer class which implements a predict method which is called by the RESTful Serving Engine. An example request is

curl -X POST http://<service-name>/predict \
    -H "Content-Type: application/json" \
    -d '{
      "data": {
        "ndarray": [[
          6,
          148,
          72,
          35,
          0,
          33.6,
          0.627,
          50
        ]],
        "names":[
          "NumTimesPrg",
          "PlGlcConc",
          "BloodP",
          "SkinThick",
          "TwoHourSerIns",
          "BMI",
          "DiPedFunc",
          "Age"
        ]
      }
    }'