README.md

# [Title]

## Introduction

This GitLab repository contains the main material used in the paper [title] by [authors].

The study examines patients undergoing treatment for alcohol disorders, utilizing machine learning techniques to predict clinical success or withdrawal. The main goal is to employ explainability tools to assess the impact of individual versus social factors on treatment outcomes. Additionally, the research explores whether the significance of these factors changed during the pandemic by comparing pre-pandemic and post-pandemic patient groups.

[Impact?]

## About the Dataset

[Origin, Characteristics]

The dataset has not been provided since the authors do not have permission for its sharing from the data providers.

### Patient Counts
| **Period**       | **Medical Discharge** | **Withdrawal**  | **Total** |
|--------------|-------------------|------------|--------|
| Pre-pandemic | 2792              | 20069      | 22861  |
| Post-pandemic| 1882              | 8795       | 10677  |
| **Overall**  | 4674              | 28864      | 33538  |

### List of Selected Features
#### Individual Factors
* Age -> Age at which the patient started the first treatment.
* Sex -> Male or Female.
* Num_Children.
* Smoking ->  Whether patient is a tobacco smoker (with abuse or dependence) at the time of starting treatment.
* Bio_Vulner ->  If the patient presents biological vulnerability, identified by the presence of 2 or more diagnoses of addictive disorders (substance use or behavioral).
* X_DXCIE ->  Evaluates the presence of diagnoses (abuse or dependence) of addictive disorders, assessed according to the ICD (International Classification of Diseases).
    * Opiods_DXCIE.
    * Cannabis_DXCIE.
    * BZD_DXCIE.
    * Cocaine_DXCIE.
    * Hallucin_DXCIE.
    * Tobacco_DXCIE.
* Frequency ->  Frequency with which patients have used their primary drug during the 30 days prior to starting treatment. One-hot encoded into:
    * Freq_1dpw.
    * Freq_2-3dpw.
    * Freq_4-6dpw.
    * Freq_l1dpw.
    * Freq_None.
    * Freq_Everyday.
* Years_Drug_Use -> Number of years the patient has been using their primary drug.
* Other_Psychiatric_DX -> Patients who are diagnosed with other psychiatric disorders.
* Previous_Treatments -> Whether patients have undergone treatments prior to the one currently being analyzed.
* Treatment_Adherence -> Ratio of the number of attended appointments to the number of scheduled appointments associated with the patient's treatment.


#### Social Factors
* Education -> Maximum education level achieved by the patient. One-hot encoded into:
    * Ed_Not_Complete_Primary
    * Ed_Primary
    * Ed_Secondary
    * Ed_Secondary_Technical
    * Ed_Tertiary
* Social_Protection -> Whether patients need or currently have social benefits.
* Job Stability -> Employment status. One-hot encoded into:
    * JobIn_Unstable
    * JobIn_Stable
    * JobIn_Unemployed
* Housing -> Housing situation. One-hot encoded into:
    * Hous_Institutional
    * Hous_Stable
    * Hous_Unstable
* Early_Alterations -> Whether patients have used drugs before the age of 11.
* Social Inclusion -> Living Situation. One-hot encoded into:
    * SocInc_Family_Friends
    * SocInc_Alone
    * SocInc_Instit
* Risk_Stigma -> Whether patients have potentially suffered social discrimination based on their status of having AIDS, hepatitis C, being injection drug users, or being women.
* Structural_Conflict -> Household income in the area where the outpatient treatment center is located.

## Dealing with Class Imbalance

One of the primary challenges we encountered was a significant class imbalance, with a higher number of patients withdrawing from treatment compared to those staying.

To address this issue, we implemented four different training approaches or pipelines on both the pre-pandemic and post-pandemic training datasets:

1. **Using the Original Dataset (ORIG)**: The models were trained on the original datasets.
2. **Class Weight Adjustment (ORIG_CW)**: The models were trained on the original datasets but were penalized more heavily for misclassifying the minority class.
3. **Oversampling (OVER)**: Additional samples were generated for the minority class (patients staying) to balance the dataset.
4. **Undersampling (UNDER)**: Samples from the majority class (patients withdrawing) were reduced to achieve balance.

These approaches resulted in multiple training datasets. However, to ensure a fair comparison of the models' performance across different pipelines, we utilized a common test dataset for evaluation, irrespective of the training approach followed.

## Methodology Overview

![repo_summary](/uploads/423d2236327507c5676b3cb497ee5285/repo_summary.png)

## Repository

This repository is organized as follows:

* [01-EDA](./01-EDA):
  * [EDA.ipynb](./01-EDA/EDA.ipynb): Exploring and filtering data, handling missing values, encoding variables, building the final pre- and post-pandemic datasets, and generating plots for feature distributions, correlations and importance.
  * [results](./01-EDA/results): 
    * [feature_names](./01-EDA/results/feature_names): Names of the selected individual and social variables.
    * [plots](./01-EDA/results/plots)
        * [correlations](./01-EDA/results/plots/correlations): Heatmaps to visualize the pairwise correlations beetween features.
        * [distributions](./01-EDA/results/plots/distributions): Statistical plots to visualize the distribution of features.
        * [feature_importance](./01-EDA/results/plots/feature_importance): Plots to show the importance of each feature in predicting outcomes.
* [02-training_data_generation](./02-training_data_generation):
  * [training_data_generation.ipynb](./02-training_data_generation/training_data_generation.ipynb): Generating training and testing datasets for each of the pipelines.
* [03-model_building](./03-model_building):
  * [hyperparameter_tuning.py](./03-model_building/hyperparameter_tuning.py): Tuning models through a random search of hyperparameters.
  * [cv_metric_generation.py](./03-model_building/cv_metric_generation.py): Generating cross-validation metrics and plots for each of the tuned models.
  * [cv_metric_distribution.py](./03-model_building/cv_metric_distribution.py): Generating boxplots for each cross-validation metric and tuned model.
  * [models_testing.py](./03-model_building/models_testing.py): Testing tuned models with test dataset.
  * [models_final_fitting.py](./03-model_building/models_final_fitting.py): Saving fitted model for each selected final model.
  * [results](./03-model_building/results):
    * [hyperparam](./03-model_building/output/hyperparam): Excel file containing the optimal hyperparameters for each model in each pipeline.
    * [cv_metrics](./03-model_building/output/cv_metrics): Material related to the results of cross-validation: scores, ROC and Precision-Recall curves and boxplots for each metric.
    * [testing](./03-model_building/output/testing): Material related to the results of testing the tuned models: scores, ROC and Precision-Recall curves and confusion matrices.
    * [fitted_models](./03-model_building/output/fitted_models): Final selected trained models.
* [04-explainability](./04-explainability):
  * [shap_vals_computation.py](./04-explainability/shap_vals_computation.py): Computing SHAP values for final models.
  * [shap_inter_vals_computation.py](./04-explainability/shap_inter_vals_computation.py): Computing SHAP interaction values for final models.
  * [shap_plots.ipynb](./04-explainability/shap_plots.ipynb): Generating SHAP summary plots for the SHAP and SHAP interaction values computed. Comparing major differences between pre- and post-pandemic groups.
  * [results](./04-explainability/results):
    * [plots](./04-explainability/plots): SHAP summary and summary interaction plots as well as 
        * [shap_summary](./04-explainability/plots/shap_summary): SHAP summary plots.
        * [shap_inter_summary](./04-explainability/plots/shap_inter_summary): SHAP summary interaction plots.
        * [heatmaps_interactions](./04-explainability/plots/heatmaps_interactions): Heatmaps representing the differences in interactions between pre-pandemic and post-pandemic groups.
        
## Contact

For any inquiry you can contact:
[Contact Info]