The problem: Predict if Medical Insurance applications are high risk

Butterfly AI can assist insurance companies decide if a medical insurance application is high risk. That will help these companies improve the accuracy of the applications approval process, while reducing the cost of insurance applications through automation and reduction of human error.

The data

The base dataset used is (InsuranceClaim.csv). It includes 98000 labelled medical insurance applications.

This dataset is the altered version of an original data which is available under a CC0: Public Domain license at https://creativecommons.org/publicdomain/zero/1.0/

First group of features: Demographics & Socioeconomic
  • person_id
  • age
  • sex
  • region
  • urban_rural
  • income
  • education
  • marital_status
  • employment_status
  • household_size
  • dependents
Second group of features: Lifestyle & Habits
  • bmi
  • smoker
  • alcohol_freq
  • exercise_frequency
  • sleep_hours
  • stress_level
Third group of features: Health & Clinical
  • hypertension
  • diabetes
  • copd
  • cardiovascular
  • cancer_history
  • kidney_disease
  • liver_disease
  • arthritis
  • mental_health
  • chronic_count
  • systolic_bp
  • diastolic_bp
  • ldl
  • hba1c
Fourth group of features: Healthcare Utilization & Procedures
  • visits_last_year
  • hospitalizations_last_3yrs
  • days_hospitalized_last_3yrs
  • medication_count
  • proc_imaging
  • proc_surgery
  • proc_psycho
  • proc_consult_count
  • proc_lab
  • had_major
Fifth group of features: Insurance & Policy
  • plan_type
  • network_tier
  • deductible
  • copay
  • policy_term_years
  • policy_changes_last_2yrs
  • provider_quality
Sixth group of features, Medical Costs & Claims:
  • annual_medical_cost
  • annual_premium
  • monthly_premium
  • claims_count
  • avg_claim_amount
  • total_claims_paid
Target of Prediction (Label):

is_high_risk



Dataset creation

Use the following parameters for dataset creation:

  • number of buckets: 40

Dataset creation


Training

This is the best training attempt:

  • scaling factor: 19
  • performance threshold: 0.97

Training params

And the created champion model:

Training best


The final performance of 0.97 was achieved after few iterations of hyperparameter tuning:

Number of Buckets Scaling Factor Performance Threshold
20 19 0.80
20 19 0.95
40 19 0.95
40 19 0.97

Final result

When performing binary classifications or predictions, Butterfly AI platform’s underlying proprietary algorithms calculate the probability of certainty for a prediction outcome.

  • One label (e.g.1) will be selected when the probability is equal or above 0.5
  • and the other one (e.g. 0) will be selected when the probability is below 0.5

The closer the value is to 0 or 1, the more certain is the prediction. The probability is presented in a dedicated column in the prediction result file.

Using this unseen unlabelled data, the resulting CSV looks like this:

Prediction 2