Business Application in Hospitals
To help medical practitioners predict the chances of patient survival proactively based on multiple variables on demographic, vitals, labs results, labs blood gas, APACHE covariate, prediction, comorbidity and grouping. Based on the prediction, the treatments could be hyper-personalized.
Prediction = Death: The doctors could hyper-personalize the treatment specific to patient conditions, such as
- Advanced diagnostics to identify specific patient needs
- Tailored interventions that target the patient's unique condition and medical history
- Possible use of experimental or aggressive treatments that may offer a higher chance of survival
Prediction = Survival: The doctors could continue with the current treatment and medication, and
- Monitor the patient closely to ensure the treatment remains effective
- Adjust the treatment plan if necessary based on the patient's response
Problem Statement
The target feature is hospital_death which is a binary variable. The task is to classify this variable based on the other 84 features based on the scoring metric: Area under ROC curve and Recall (this is important as the cost and risk of model predicting death is higher than the model predicting life)
Data cleaning and preprocessing
- Outlier capping based on the APACHE3 medical standards
- Auto EDA using autoviz
Deep Learning Model Building and Evaluation using Keras
- Optuna-based Hyperparameter tuning and Kerascheckback to prune inefficient trials
- Model checkpoint to save the best model state
- Backtracking the Streamlit results to verify the model's prediction
Results
- Validation AUC of 84% and Recall of 78%
Application UI
-
In medical context, practitioners typical upload patient results. Hence, this application has the option to upload patient information (.csv).
Note: TestData_Patients is available in the Dataset folder
-
Link: https://patientsurvivaldetection-gjfgzmmltdkgycabkjnrcz.streamlit.app/
Next Steps 📃☑️✅
- Log results on MLFlow
- Showcase the results on DagsHub