health

PREPARE Challenge - Phase 2: Model Arena (Social Determinants Track)

Advance algorithms and analytic approaches for early prediction of Alzheimer's disease and related dementias based on social determinants of health, with an emphasis on explainability of predictions. [Model Arena - Social Determinants Track] #health

$130,000 in prizes
Completed dec 2024
440 joined

Advance algorithms and analytic approaches for early prediction of Alzheimer's disease and related dementias based on social determinants of health, with an emphasis on explainability of predictions. [Model Arena - Social Determinants Track]

Quick Facts

Participants
440
No. of Entries
2,091
Prize
$130,000
Winner

RASKA-Team

1st Place Team

Navigation

Quick access

Community Code

Predicting AD Feature Engineering and Modeling Starter Notebook

⸱
4mo 3w ago
⸱
96 downloads

Download this code
Code requirements (Download)
utility.py:
from sklearn.model_selection import StratifiedKFold, KFold, cross_val_score
from sklearn.metrics import roc_auc_score, accuracy_score, confusion_matrix, ConfusionMatrixDisplay, f1_score, precision_score, recall_score, mean_squared_error
import numpy as np
from sklearn.base import clone
from tqdm import tqdm
from IPython.display import clear_output
import matplotlib.pyplot as plt
from sklearn.model_selection import train_test_split
import pandas as pd
from sklearn.compose import ColumnTransformer
from sklearn.preprocessing import MinMaxScaler, StandardScaler, OneHotEncoder, LabelEncoder
from lightgbm  import early_stopping 

# create new features by combining 2003 and 2012 scores and numbering ordinal variables
def feature_engineering(data):
    data['rjob_hrswk_change']=data['rjob_hrswk_12'].fillna(0)-data['rjob_hrswk_03'].fillna(0).astype(float)
    data['rjob_hrswk']=data[['rjob_hrswk_03','rjob_hrswk_12']].mean(axis=1).astype(float)
    data['max_work_year']=data[['rjob_end_12','rjob_end_03']].max(axis=1).astype(float)
    data['years_since_work']=(data['year']-data['max_work_year']).astype(float)
    data['hincome_change']=(data['hincome_12'].fillna(0)-data['hincome_03'].fillna(0)).astype(float)
    data['niadl_change']=(data['n_iadl_12'].fillna(0)-data['n_iadl_03'].fillna(0)).astype(float)
    data['adl_max']=(data[['n_adl_12','n_adl_03']].fillna(0)).max(axis=1).astype(float)
    data['iadl_max']=(data[['n_iadl_12','n_iadl_03']].fillna(0)).max(axis=1).astype(float)
    data['neg_adl']=6-data['adl_max']
    data['neg_iadl']=4-data['iadl_max']
    data['adl_change']=(data['n_adl_12'].fillna(0)-data['n_adl_03'].fillna(0)).astype(float)
    data['depr_change']=(data['n_depr_12'].fillna(0)-data['n_depr_03'].fillna(0)).astype(float)
    data['glob_hlth_03']=data['glob_hlth_03'].replace({'5. Poor':0, '4. Fair':1, '3. Good':2, '2. Very good': 3, '1. Excellent':4}).astype(float)
    data['glob_hlth_12']=data['glob_hlth_12'].replace({'5. Poor':0, '4. Fair':1, '3. Good':2, '2. Very good': 3, '1. Excellent':4}).astype(float)
    data['glob_hlth']=data[['glob_hlth_03', 'glob_hlth_12']].sum(axis=1).astype(float)
    data['bmi_03']=data['bmi_03'].replace({'1. Underweight': 1, '2. Normal weight': 2, '3. Overweight':3, '4. Obese':4, '5. Morbidly obese':5}).astype(float)
    data['bmi_12']=data['bmi_12'].replace({'1. Underweight': 1, '2. Normal weight': 2, '3. Overweight':3, '4. Obese':4, '5. Morbidly obese':5}).astype(float)
    data['bmi']=data[['bmi_03', 'bmi_12']].sum(axis=1).astype(float)
    data['employment_03']=data['employment_03'].replace({'1. Currently Working': 'Working', '2. Currently looking for work':'Looking for work', '3. Dedicated to household chores': 'House', '4. Retired, incapacitated, or does not work': 'No work'})
    data['employment_12']=data['employment_12'].replace({'1. Currently Working': 'Working', '2. Currently looking for work':'Looking for work', '3. Dedicated to household chores': 'House', '4. Retired, incapacitated, or does not work': 'No work'})
    data['memory_12']=data['memory_12'].replace({'5. Poor':0, '4. Fair':1, '3. Good':2, '2. Very good': 3, '1. Excellent':4}).astype(float)
    data['edu_gru_03']=data['edu_gru_03'].replace({'0. No education':0,'1. 1–5 years':1, '2. 6 years':2, '3. 7–9 years':3,'4. 10+ years':4}).astype(float)
    data['edu_gru_12']=data['edu_gru_12'].replace({'0. No education':0,'1. 1–5 years':1, '2. 6 years':2, '3. 7–9 years':3,'4. 10+ years':4}).astype(float)
    data['urban_03']=data['urban_03'].fillna(0).replace({ '1. 100,000+':2, '0. <100,000':1}).astype(float)
    data['urban_12']=data['urban_03'].fillna(0).replace({ '1. 100,000+':2, '0. <100,000':1}).astype(float)
    data['issste']=data[['issste_03', 'issste_12']].fillna(0).max(axis=1).astype(float)
    data['urban']=data[['urban_03', 'urban_12']].max(axis=1).astype(float)
    data['edu_gru']=data[['edu_gru_03', 'edu_gru_12']].max(axis=1).astype(float)
    data['hincome']=data[['hincome_03', 'hincome_12']].max(axis=1).astype(float)
    data['illnesses']=data[['n_illnesses_03', 'n_illnesses_12']].max(axis=1).astype(float)
    data['alc_tob_03']=data[['alcohol_03','tobacco_03']].sum(axis=1).astype(float)
    data['alc_tob_12']=data[['alcohol_12','tobacco_12']].sum(axis=1).astype(float)
    data['alc_tob']=data[['alc_tob_03', 'alc_tob_12']].max(axis=1).astype(float)
    data['rearnings']=data[['rearnings_03', 'rearnings_12']].max(axis=1).astype(float)
    data.drop(columns=['issste_03', 'issste_12','urban_03', 'urban_12','edu_gru_03', 'edu_gru_12','bmi_03', 'bmi_12','alc_tob_03', 'alc_tob_12','alcohol_03','tobacco_03','alcohol_12','tobacco_12','hincome_03', 'hincome_12' ], inplace=True)
    return data

#get final features for model based on correlation with target variable and removing variables that are highly correlated with each other
def get_final_features(train_data,threshold):

    features=[]

    # remove ID, other identifying/non-input columns
    for col in train_data.columns:
        if 'uid' not in col:
            features.append(col)

    #select final features based on correlation with target variable
    correlations=train_data[features].corr()[['composite_score']]
    final_features=list(correlations[((correlations['composite_score']>.07) | (correlations['composite_score']<-.07))].T.columns.values) 

    #remove other redundant features
    threshold = threshold
    correlation_matrix = train_data[final_features].drop(columns='composite_score').corr()
    highly_correlated_features = set()
    for i in range(len(correlation_matrix.columns)):
        for j in range(i):
            if abs(correlation_matrix.iloc[i, j]) > threshold:
                colname = correlation_matrix.columns[i]
                highly_correlated_features.add(colname)

    final_features=list(set(final_features)-highly_correlated_features)

    return final_features

#train model with Kfolds based on the unique uids and return the RMSE on t he validation set
def TrainML(model_class, df ,final_features, n_splits,SEED):
    SKF = KFold(n_splits=n_splits, shuffle=True, random_state=SEED)
    train_S = []
    test_S = []
    oof_non_rounded = np.zeros(len(df), dtype=float) 
    unique_uids = df['uid'].unique()
    for fold, (train_id, val_id) in enumerate(tqdm(SKF.split(unique_uids), desc="Training Folds", total=n_splits)):
        df_train, df_val = df.iloc[train_id], df.iloc[val_id]
        X_train, X_val = df_train[final_features], df_val[final_features]
        y_train, y_val=df_train['composite_score'], df_val['composite_score']
        model = clone(model_class)
        model.fit(X_train, y_train)
        y_train_pred = model.predict(X_train)
        y_val_pred = model.predict(X_val)
        oof_non_rounded[val_id] = y_val_pred
        train_rmse = mean_squared_error(y_train, y_train_pred.round(0).astype(int), squared=False)
        val_rmse = mean_squared_error(y_val, y_val_pred.round(0).astype(int), squared=False)
        train_S.append(train_rmse)
        test_S.append(val_rmse)
        print(f"Fold {fold+1} - Train AUC: {train_rmse:.4f}, Validation AUC: {val_rmse:.4f}")
        clear_output(wait=True)

    print(f"Mean Train RMSE --> {np.mean(train_S):.4f}")
    print(f"Mean Validation RMSE ---> {np.mean(test_S):.4f}")




def get_cat_cols(data):
    # Get the columns with object datatype
    cat_columns=[]
    dummies=[]
    for col in data.columns:
        if data[col].dtype=='object' and 'uid' not in col:
            cat_columns.append(col)
            dummies.append(col)
        elif data[col].dtype!='object' and 'uid' not in col and (data[col].max()==1.0):
            cat_columns.append(col)
            data[col].fillna(0, inplace=True)
        else:
            continue
    return cat_columns, dummies


#encode categorical columns with OneHotEncoder
def encode_cat_cols(train_data, test_data, cat_cols, dummy_cols):
    enc = OneHotEncoder()
    enc.fit(train_data[dummy_cols])
    encoded_train_data=enc.transform(train_data[dummy_cols]).toarray()
    encoded_test_data=enc.transform(test_data[dummy_cols]).toarray()
    feature_names = enc.get_feature_names_out(dummy_cols)
    train_data.drop(columns=dummy_cols, inplace=True)
    test_data.drop(columns=dummy_cols, inplace=True)
    encoded_train_df = pd.DataFrame(encoded_train_data, columns=feature_names)
    encoded_test_df = pd.DataFrame(encoded_test_data, columns=feature_names)
    train_data[feature_names]=encoded_train_df[feature_names]
    test_data[feature_names]=encoded_test_df[feature_names]
    return train_data, test_data, feature_names
Code version history (1)
v1
⸱
4mo 3w ago
⸱
Latest

Tags

#python #Preprocessing #Modeling #feature-selection

License

MIT open source