Predicting Chronic Heart Failure - Solution

Predicting Chronic Heart Failure - Solution#

Spoiler:
This notebook contains a comprehensive step-by-step solution to the exercise.
Try working through the exercise on your own first before looking at this suggested solution.

Clinical Use Case#

Patients admitted with myocardial infarction (MI) are at risk of developing a range of complications, including chronic heart failure (CHF).

Chronic heart failure is a serious condition that can significantly impair quality of life and is associated with increased morbidity and mortality.

Following an acute myocardial infarction, some patients recover without long-term consequences, while others develop progressive cardiac dysfunction leading to heart failure.

Early identification of patients at risk of developing chronic heart failure is challenging, even for experienced clinicians, but highly relevant for optimizing treatment and improving long-term outcomes.

Goal of this analysis:

Build a machine learning model that predicts whether a patient will develop chronic heart failure during hospitalization.

You can download the dataset of myocardial infarction complications from the University of Leicester here: https://figshare.le.ac.uk/ndownloader/files/23581310

About the Dataset#

This dataset contains clinical information about patients admitted with myocardial infarction and was designed to evaluate real-world medical prediction problems.

Variables include:

  • demographic data

  • medical history

  • ECG findings

  • laboratory values

  • treatment information

Possible complications are stored in the target variables.

In this notebook, we focus on predicting:

Chronic Heart Failure

Additional information about the dataset, including variable descriptions, can be found here: https://doi.org/10.25392/leicester.data.12045261

Important methodological aspect

The dataset allows prediction at different time points during the hospital stay:

  1. At admission

  2. After 24 hours

  3. After 48 hours

  4. After 72 hours

Depending on the chosen time point, different variables are available.

For this exercise, you must decide on one time point and adapt your feature selection accordingly.

For example:

  • If you predict at admission, you may only use variables available at admission

  • Later time points allow more information, but also introduce the risk of data leakage

This reflects a key challenge in clinical machine learning:

Predictions must be based only on information that is available at the time the decision is made.

Potential clinical use:

  • early identification of patients at risk of chronic heart failure

  • timely initiation of preventive or therapeutic interventions

  • improved long-term management and follow-up planning

Your Tasks#

  • Load and explore the dataset to understand its structure and contents

  • Decide at which time point you want to predict the ventricular fibrillation (target variable = “FIBR_JELUD”)
    Adjust your feature selection accordingly

  • Prepare the data for machine learning

  • Train and compare different models (e.g. Logistic Regression, Random Forest, XGB)

  • Evaluate model performance using appropriate metrics

  • Interpret your results and reflect on their clinical relevance

# Import bia-bob as a helpful Python & Medical AI expert
from bia_bob import bob
import os

bob.initialize(
    endpoint=os.getenv('ENDPOINT_URL'),
    model="vllm-llama-4-scout-17b-16e-instruct",
    system_prompt=os.getenv('SYSTEM_PROMPT_MEDICAL_AI')
)

%bob Who are you? Just one sentence!

NOTE#

In this notebook we focus on predicting:

Chronic heart failure (ZSN) at the time of admission to the hospital

Therefore, we use all input columns (2-112) except 93, 94, 95, 100, 101, 102, 103, 104, 105.

Step 1: Load and inspect the data#

import pandas as pd
import numpy as np

import matplotlib.pyplot as plt
import seaborn as sns

from sklearn.model_selection import train_test_split
from sklearn.model_selection import GridSearchCV

from sklearn.pipeline import Pipeline
from sklearn.compose import ColumnTransformer

from sklearn.preprocessing import StandardScaler
from sklearn.preprocessing import OneHotEncoder

from sklearn.impute import SimpleImputer

from sklearn.linear_model import LogisticRegression
from sklearn.ensemble import RandomForestClassifier

from xgboost import XGBClassifier

from sklearn.metrics import roc_auc_score
from sklearn.metrics import classification_report
from sklearn.metrics import roc_curve
from sklearn.metrics import confusion_matrix
from sklearn.metrics import RocCurveDisplay

import shap

pd.set_option('display.max_rows', 500)
pd.set_option('display.max_columns', 500)

# load the dataset
df = pd.read_csv('Myocardial_infarction_complications_Database.csv')

df.head()
ID AGE SEX INF_ANAM STENOK_AN FK_STENOK IBS_POST IBS_NASL GB SIM_GIPERT DLIT_AG ZSN_A nr_11 nr_01 nr_02 nr_03 nr_04 nr_07 nr_08 np_01 np_04 np_05 np_07 np_08 np_09 np_10 endocr_01 endocr_02 endocr_03 zab_leg_01 zab_leg_02 zab_leg_03 zab_leg_04 zab_leg_06 S_AD_KBRIG D_AD_KBRIG S_AD_ORIT D_AD_ORIT O_L_POST K_SH_POST MP_TP_POST SVT_POST GT_POST FIB_G_POST ant_im lat_im inf_im post_im IM_PG_P ritm_ecg_p_01 ritm_ecg_p_02 ritm_ecg_p_04 ritm_ecg_p_06 ritm_ecg_p_07 ritm_ecg_p_08 n_r_ecg_p_01 n_r_ecg_p_02 n_r_ecg_p_03 n_r_ecg_p_04 n_r_ecg_p_05 n_r_ecg_p_06 n_r_ecg_p_08 n_r_ecg_p_09 n_r_ecg_p_10 n_p_ecg_p_01 n_p_ecg_p_03 n_p_ecg_p_04 n_p_ecg_p_05 n_p_ecg_p_06 n_p_ecg_p_07 n_p_ecg_p_08 n_p_ecg_p_09 n_p_ecg_p_10 n_p_ecg_p_11 n_p_ecg_p_12 fibr_ter_01 fibr_ter_02 fibr_ter_03 fibr_ter_05 fibr_ter_06 fibr_ter_07 fibr_ter_08 GIPO_K K_BLOOD GIPER_NA NA_BLOOD ALT_BLOOD AST_BLOOD KFK_BLOOD L_BLOOD ROE TIME_B_S R_AB_1_n R_AB_2_n R_AB_3_n NA_KB NOT_NA_KB LID_KB NITR_S NA_R_1_n NA_R_2_n NA_R_3_n NOT_NA_1_n NOT_NA_2_n NOT_NA_3_n LID_S_n B_BLOK_S_n ANT_CA_S_n GEPAR_S_n ASP_S_n TIKL_S_n TRENT_S_n FIBR_PREDS PREDS_TAH JELUD_TAH FIBR_JELUD A_V_BLOK OTEK_LANC RAZRIV DRESSLER ZSN REC_IM P_IM_STEN LET_IS
0 1 77.0 1 2.0 1.0 1.0 2.0 NaN 3.0 0.0 7.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 NaN NaN 180.0 100.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 4.7 0.0 138.0 NaN NaN NaN 8.0 16.0 4.0 0.0 0.0 1.0 NaN NaN NaN 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 1.0 1.0 0.0 0.0 0 0 0 0 0 0 0 0 0 0 0 0
1 2 55.0 1 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 NaN NaN 120.0 90.0 0.0 0.0 0.0 0.0 0.0 0.0 4.0 1.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 3.5 0.0 132.0 0.38 0.18 NaN 7.8 3.0 2.0 0.0 0.0 0.0 1.0 0.0 1.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 1.0 0.0 1.0 1.0 1.0 0.0 1.0 0 0 0 0 0 0 0 0 0 0 0 0
2 3 52.0 1 0.0 0.0 0.0 2.0 NaN 2.0 0.0 2.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 150.0 100.0 180.0 100.0 0.0 0.0 0.0 0.0 0.0 0.0 4.0 1.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 4.0 0.0 132.0 0.30 0.11 NaN 10.8 NaN 3.0 3.0 0.0 0.0 1.0 1.0 1.0 0.0 1.0 0.0 0.0 3.0 2.0 2.0 1.0 1.0 0.0 1.0 1.0 0.0 0.0 0 0 0 0 0 0 0 0 0 0 0 0
3 4 68.0 0 0.0 0.0 0.0 2.0 NaN 2.0 0.0 3.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 NaN NaN 120.0 70.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 1.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 3.9 0.0 146.0 0.75 0.37 NaN NaN NaN 2.0 0.0 0.0 1.0 NaN NaN NaN 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 1.0 1.0 0.0 0.0 0 0 0 0 0 0 0 0 1 0 0 0
4 5 60.0 1 0.0 0.0 0.0 2.0 NaN 3.0 0.0 7.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 190.0 100.0 160.0 90.0 0.0 0.0 0.0 0.0 0.0 0.0 4.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 3.5 0.0 132.0 0.45 0.22 NaN 8.3 NaN 9.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 1.0 0.0 1.0 0 0 0 0 0 0 0 0 0 0 0 0 
# Get dataset size
print("Dataset contains {} rows".format(df.shape[0]))
print("Dataset contains {} columns".format(df.shape[1]))

Dataset contains 1700 rows
Dataset contains 124 columns

# Get summary
df.info()

<class 'pandas.core.frame.DataFrame'>
RangeIndex: 1700 entries, 0 to 1699
Columns: 124 entries, ID to LET_IS
dtypes: float64(110), int64(14)
memory usage: 1.6 MB

# Set patient ID as index (keep column as well)
df = df.set_index("ID", drop=False)

# Define target variable
target = "ZSN"

# Feature Selection (Admission)

# Select predictor columns (2–112)
feature_cols = df.columns[1:112]

# Indices to remove (relative to original df)
drop_idx = [92, 93, 94, 99, 100, 101, 102, 103, 104]

cols_to_remove = df.columns[drop_idx]

# Final feature list
feature_cols = feature_cols.drop(cols_to_remove)

# Keep only relevant columns in df
# IMPORTANT: keep target + selected features
df = df[feature_cols.tolist() + [target]]

# Optional: rename target for clarity
df = df.rename(columns={target: "target"})

# Get new size of the dataset
df.shape

(1700, 103)

df.head()
AGE SEX INF_ANAM STENOK_AN FK_STENOK IBS_POST IBS_NASL GB SIM_GIPERT DLIT_AG ZSN_A nr_11 nr_01 nr_02 nr_03 nr_04 nr_07 nr_08 np_01 np_04 np_05 np_07 np_08 np_09 np_10 endocr_01 endocr_02 endocr_03 zab_leg_01 zab_leg_02 zab_leg_03 zab_leg_04 zab_leg_06 S_AD_KBRIG D_AD_KBRIG S_AD_ORIT D_AD_ORIT O_L_POST K_SH_POST MP_TP_POST SVT_POST GT_POST FIB_G_POST ant_im lat_im inf_im post_im IM_PG_P ritm_ecg_p_01 ritm_ecg_p_02 ritm_ecg_p_04 ritm_ecg_p_06 ritm_ecg_p_07 ritm_ecg_p_08 n_r_ecg_p_01 n_r_ecg_p_02 n_r_ecg_p_03 n_r_ecg_p_04 n_r_ecg_p_05 n_r_ecg_p_06 n_r_ecg_p_08 n_r_ecg_p_09 n_r_ecg_p_10 n_p_ecg_p_01 n_p_ecg_p_03 n_p_ecg_p_04 n_p_ecg_p_05 n_p_ecg_p_06 n_p_ecg_p_07 n_p_ecg_p_08 n_p_ecg_p_09 n_p_ecg_p_10 n_p_ecg_p_11 n_p_ecg_p_12 fibr_ter_01 fibr_ter_02 fibr_ter_03 fibr_ter_05 fibr_ter_06 fibr_ter_07 fibr_ter_08 GIPO_K K_BLOOD GIPER_NA NA_BLOOD ALT_BLOOD AST_BLOOD KFK_BLOOD L_BLOOD ROE TIME_B_S NA_KB NOT_NA_KB LID_KB NITR_S LID_S_n B_BLOK_S_n ANT_CA_S_n GEPAR_S_n ASP_S_n TIKL_S_n TRENT_S_n target
ID
1 77.0 1 2.0 1.0 1.0 2.0 NaN 3.0 0.0 7.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 NaN NaN 180.0 100.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 4.7 0.0 138.0 NaN NaN NaN 8.0 16.0 4.0 NaN NaN NaN 0.0 1.0 0.0 0.0 1.0 1.0 0.0 0.0 0
2 55.0 1 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 NaN NaN 120.0 90.0 0.0 0.0 0.0 0.0 0.0 0.0 4.0 1.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 3.5 0.0 132.0 0.38 0.18 NaN 7.8 3.0 2.0 1.0 0.0 1.0 0.0 1.0 0.0 1.0 1.0 1.0 0.0 1.0 0
3 52.0 1 0.0 0.0 0.0 2.0 NaN 2.0 0.0 2.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 150.0 100.0 180.0 100.0 0.0 0.0 0.0 0.0 0.0 0.0 4.0 1.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 4.0 0.0 132.0 0.30 0.11 NaN 10.8 NaN 3.0 1.0 1.0 1.0 0.0 1.0 1.0 0.0 1.0 1.0 0.0 0.0 0
4 68.0 0 0.0 0.0 0.0 2.0 NaN 2.0 0.0 3.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 NaN NaN 120.0 70.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 1.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 3.9 0.0 146.0 0.75 0.37 NaN NaN NaN 2.0 NaN NaN NaN 0.0 0.0 0.0 1.0 1.0 1.0 0.0 0.0 1
5 60.0 1 0.0 0.0 0.0 2.0 NaN 3.0 0.0 7.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 190.0 100.0 160.0 90.0 0.0 0.0 0.0 0.0 0.0 0.0 4.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 3.5 0.0 132.0 0.45 0.22 NaN 8.3 NaN 9.0 0.0 0.0 0.0 0.0 0.0 0.0 1.0 0.0 1.0 0.0 1.0 0 
# Detect feature types

binary_features = []
categorical_features = []
continuous_features = []

for col in feature_cols:

    n_unique = df[col].nunique(dropna=True)

    # Binary: exactly 2 unique values
    if n_unique == 2:
        binary_features.append(col)

    # Categorical: few unique values (but not binary)
    elif n_unique < 10:
        categorical_features.append(col)

    # Continuous: many unique values
    else:
        continuous_features.append(col)

# Assign data types
df[binary_features] = df[binary_features].astype("category")
df[categorical_features] = df[categorical_features].astype("category")

# Overview of 
print("Binary features:", len(binary_features))
print("Categorical features:", len(categorical_features))
print("Continuous features:", len(continuous_features))

Binary features: 78
Categorical features: 13
Continuous features: 11

df.dtypes.value_counts()

category    77
float64     11
category     6
category     2
category     1
category     1
category     1
category     1
category     1
category     1
int64        1
Name: count, dtype: int64

# Summary statistics for numeric columns
df.describe(include=['int', 'float']).T
count mean std min 25% 50% 75% max
AGE 1692.0 61.856974 11.259936 26.00 54.00 63.00 70.00 92.00
S_AD_KBRIG 624.0 136.907051 34.997835 0.00 120.00 140.00 160.00 260.00
D_AD_KBRIG 624.0 81.394231 19.745045 0.00 70.00 80.00 90.00 190.00
S_AD_ORIT 1433.0 134.588276 31.348388 0.00 120.00 130.00 150.00 260.00
D_AD_ORIT 1433.0 82.749477 18.321063 0.00 80.00 80.00 90.00 190.00
K_BLOOD 1329.0 4.191422 0.754076 2.30 3.70 4.10 4.60 8.20
NA_BLOOD 1325.0 136.550943 6.512120 117.00 133.00 136.00 140.00 169.00
ALT_BLOOD 1416.0 0.481455 0.387261 0.03 0.23 0.38 0.61 3.00
AST_BLOOD 1415.0 0.263717 0.201802 0.04 0.15 0.22 0.33 2.15
L_BLOOD 1575.0 8.782914 3.400557 2.00 6.40 8.00 10.45 27.90
ROE 1497.0 13.444890 11.296316 1.00 5.00 10.00 18.00 140.00
target 1700.0 0.231765 0.422084 0.00 0.00 0.00 0.00 1.00 
# Summary statistics for categorical columns
df.describe(include="category").T
count unique top freq
SEX 1700.0 2.0 1.0 1065.0
INF_ANAM 1696.0 4.0 0.0 1060.0
STENOK_AN 1594.0 7.0 0.0 661.0
FK_STENOK 1627.0 5.0 2.0 854.0
IBS_POST 1649.0 3.0 2.0 683.0
IBS_NASL 72.0 2.0 0.0 45.0
GB 1691.0 4.0 2.0 880.0
SIM_GIPERT 1692.0 2.0 0.0 1635.0
DLIT_AG 1452.0 8.0 0.0 551.0
ZSN_A 1646.0 5.0 0.0 1468.0
nr_11 1679.0 2.0 0.0 1637.0
nr_01 1679.0 2.0 0.0 1675.0
nr_02 1679.0 2.0 0.0 1660.0
nr_03 1679.0 2.0 0.0 1644.0
nr_04 1679.0 2.0 0.0 1650.0
nr_07 1679.0 2.0 0.0 1678.0
nr_08 1679.0 2.0 0.0 1675.0
np_01 1682.0 2.0 0.0 1680.0
np_04 1682.0 2.0 0.0 1679.0
np_05 1682.0 2.0 0.0 1671.0
np_07 1682.0 2.0 0.0 1681.0
np_08 1682.0 2.0 0.0 1676.0
np_09 1682.0 2.0 0.0 1680.0
np_10 1682.0 2.0 0.0 1679.0
endocr_01 1689.0 2.0 0.0 1461.0
endocr_02 1690.0 2.0 0.0 1648.0
endocr_03 1690.0 2.0 0.0 1677.0
zab_leg_01 1693.0 2.0 0.0 1559.0
zab_leg_02 1693.0 2.0 0.0 1572.0
zab_leg_03 1693.0 2.0 0.0 1656.0
zab_leg_04 1693.0 2.0 0.0 1684.0
zab_leg_06 1693.0 2.0 0.0 1671.0
O_L_POST 1688.0 2.0 0.0 1578.0
K_SH_POST 1685.0 2.0 0.0 1639.0
MP_TP_POST 1686.0 2.0 0.0 1572.0
SVT_POST 1688.0 2.0 0.0 1680.0
GT_POST 1688.0 2.0 0.0 1680.0
FIB_G_POST 1688.0 2.0 0.0 1673.0
ant_im 1617.0 5.0 0.0 660.0
lat_im 1620.0 5.0 1.0 838.0
inf_im 1620.0 5.0 0.0 937.0
post_im 1628.0 5.0 0.0 1370.0
IM_PG_P 1699.0 2.0 0.0 1649.0
ritm_ecg_p_01 1548.0 2.0 1.0 1029.0
ritm_ecg_p_02 1548.0 2.0 0.0 1453.0
ritm_ecg_p_04 1548.0 2.0 0.0 1525.0
ritm_ecg_p_06 1548.0 2.0 0.0 1547.0
ritm_ecg_p_07 1548.0 2.0 0.0 1195.0
ritm_ecg_p_08 1548.0 2.0 0.0 1502.0
n_r_ecg_p_01 1585.0 2.0 0.0 1527.0
n_r_ecg_p_02 1585.0 2.0 0.0 1577.0
n_r_ecg_p_03 1585.0 2.0 0.0 1381.0
n_r_ecg_p_04 1585.0 2.0 0.0 1516.0
n_r_ecg_p_05 1585.0 2.0 0.0 1515.0
n_r_ecg_p_06 1585.0 2.0 0.0 1553.0
n_r_ecg_p_08 1585.0 2.0 0.0 1581.0
n_r_ecg_p_09 1585.0 2.0 0.0 1583.0
n_r_ecg_p_10 1585.0 2.0 0.0 1583.0
n_p_ecg_p_01 1585.0 2.0 0.0 1583.0
n_p_ecg_p_03 1585.0 2.0 0.0 1553.0
n_p_ecg_p_04 1585.0 2.0 0.0 1580.0
n_p_ecg_p_05 1585.0 2.0 0.0 1583.0
n_p_ecg_p_06 1585.0 2.0 0.0 1558.0
n_p_ecg_p_07 1585.0 2.0 0.0 1483.0
n_p_ecg_p_08 1585.0 2.0 0.0 1578.0
n_p_ecg_p_09 1585.0 2.0 0.0 1575.0
n_p_ecg_p_10 1585.0 2.0 0.0 1551.0
n_p_ecg_p_11 1585.0 2.0 0.0 1557.0
n_p_ecg_p_12 1585.0 2.0 0.0 1507.0
fibr_ter_01 1690.0 2.0 0.0 1677.0
fibr_ter_02 1690.0 2.0 0.0 1674.0
fibr_ter_03 1690.0 2.0 0.0 1622.0
fibr_ter_05 1690.0 2.0 0.0 1686.0
fibr_ter_06 1690.0 2.0 0.0 1681.0
fibr_ter_07 1690.0 2.0 0.0 1684.0
fibr_ter_08 1690.0 2.0 0.0 1688.0
GIPO_K 1331.0 2.0 0.0 797.0
GIPER_NA 1325.0 2.0 0.0 1295.0
KFK_BLOOD 4.0 4.0 1.2 1.0
TIME_B_S 1574.0 9.0 2.0 360.0
NA_KB 1043.0 2.0 1.0 618.0
NOT_NA_KB 1014.0 2.0 1.0 701.0
LID_KB 1023.0 2.0 0.0 627.0
NITR_S 1691.0 2.0 0.0 1496.0
LID_S_n 1690.0 2.0 0.0 1211.0
B_BLOK_S_n 1689.0 2.0 0.0 1474.0
ANT_CA_S_n 1687.0 2.0 1.0 1125.0
GEPAR_S_n 1683.0 2.0 1.0 1203.0
ASP_S_n 1683.0 2.0 1.0 1252.0
TIKL_S_n 1684.0 2.0 0.0 1654.0
TRENT_S_n 1684.0 2.0 0.0 1343.0

Step 2: Exploratory Data Analysis (EDA)#

Before training machine learning models, it is important to understand the structure of the dataset.

In this section we explore:

  • distribution of target variable

  • missing values

  • class imbalance

df['target'].value_counts()

target
0    1306
1     394
Name: count, dtype: int64

sns.countplot(x='target', data=df)
plt.title("Distribution of Chronic Heart Failure")
plt.show()
../_images/490c5b190f9e476a13e94a4f6e8e93f80320a866bcbc78046eb399e11117d6c8.png

-> Class Imbalance

Chronic Heart Failure is a rare event in this dataset.

This creates a class imbalance problem, meaning that most patients belong to the negative class.

This can lead to biased models that simply predict the majority class.

To address this issue we will later use:

  • class weights

  • adjusted thresholds

  • ROC-AUC evaluation

Visualization of prediction variables#

# NOTE: 0 - female, 1 - male
sns.countplot(x=df['SEX'].map({0: "Female", 1: "Male"}), data=df)
plt.title("Sex Distribution")
plt.show()
../_images/601b4e4822a108a5c97bc7c07b9fdfaecfbd70fb808f251540d0b0852b969ca4.png 
sns.boxplot(x=df['target'].map({0: "No CHF", 1: "CHF"}), y="AGE", data=df)
plt.title("Age vs Chronic Heart Failure")
plt.show()
../_images/d47fedc6c2287b6af622492f1a8ce5738c63215b75153fecb97f49183f5da8e2.png 
import matplotlib.pyplot as plt
import math

# number of plots
n_features = len(continuous_features)

# define grid size
n_cols = 3
n_rows = math.ceil(n_features / n_cols)

# create figure
fig, axes = plt.subplots(n_rows, n_cols, figsize=(15, 4 * n_rows))

# flatten axes (important!)
axes = axes.flatten()

# plot each feature
for i, col in enumerate(continuous_features):

    axes[i].hist(df[col].dropna(), bins=30)
    axes[i].set_title(col)

# remove empty plots
for j in range(i + 1, len(axes)):
    fig.delaxes(axes[j])

plt.tight_layout()
plt.show()
../_images/eeb42eb665e41677c16fe25c9d74ec51201334fd9ceff37b46cb6338c5c006f7.png 
# number of plots
n_features = len(categorical_features)

# grid size
n_cols = 3
n_rows = math.ceil(n_features / n_cols)

# create figure
fig, axes = plt.subplots(n_rows, n_cols, figsize=(15, 4 * n_rows))

axes = axes.flatten()

# plot each categorical feature
for i, col in enumerate(categorical_features):

    sns.countplot(x=df[col], ax=axes[i])
    axes[i].set_title(col)
    axes[i].tick_params(axis='x', rotation=45)

# remove empty plots
for j in range(i + 1, len(axes)):
    fig.delaxes(axes[j])

plt.tight_layout()
plt.show()
../_images/d5efa76f019ab7d3a42da7bd0c674842946470139354ec42e62ce43b6caa0981.png

Check for missing values#

# Get table with missing value counts and percentage for each feature
missing_summary = pd.DataFrame({
    "missing_count": df.isna().sum(),
    "missing_percent": df.isna().mean() * 100
})

missing_summary = missing_summary.sort_values(
    "missing_percent", ascending=False
)

missing_summary.head(25)
missing_count missing_percent
KFK_BLOOD 1696 99.764706
IBS_NASL 1628 95.764706
D_AD_KBRIG 1076 63.294118
S_AD_KBRIG 1076 63.294118
NOT_NA_KB 686 40.352941
LID_KB 677 39.823529
NA_KB 657 38.647059
NA_BLOOD 375 22.058824
GIPER_NA 375 22.058824
K_BLOOD 371 21.823529
GIPO_K 369 21.705882
AST_BLOOD 285 16.764706
ALT_BLOOD 284 16.705882
D_AD_ORIT 267 15.705882
S_AD_ORIT 267 15.705882
DLIT_AG 248 14.588235
ROE 203 11.941176
ritm_ecg_p_08 152 8.941176
ritm_ecg_p_07 152 8.941176
ritm_ecg_p_04 152 8.941176
ritm_ecg_p_02 152 8.941176
ritm_ecg_p_01 152 8.941176
ritm_ecg_p_06 152 8.941176
TIME_B_S 126 7.411765
L_BLOOD 125 7.352941

Correlation Analysis#

We inspect correlations between continuous variables.

Highly correlated variables often measure similar physiological processes.

Tree-based models such as Random Forest and XGBoost can handle correlated predictors relatively well, therefore we do not automatically remove them.

# Calculate correlation matrix for numerical features
corr_matrix = df.select_dtypes(include=[np.number]).corr()

# Show correlation table
corr_matrix
AGE S_AD_KBRIG D_AD_KBRIG S_AD_ORIT D_AD_ORIT K_BLOOD NA_BLOOD ALT_BLOOD AST_BLOOD L_BLOOD ROE target
AGE 1.000000 0.095658 -0.022013 0.043821 -0.049489 -0.002999 0.031139 -0.104688 -0.053533 0.003120 0.214393 0.146107
S_AD_KBRIG 0.095658 1.000000 0.844144 0.611365 0.555501 -0.004314 -0.008899 -0.045470 -0.083252 -0.172723 0.005249 0.072490
D_AD_KBRIG -0.022013 0.844144 1.000000 0.543048 0.555960 -0.011671 0.012838 -0.056683 -0.057641 -0.125900 -0.022010 0.059376
S_AD_ORIT 0.043821 0.611365 0.543048 1.000000 0.861266 0.030007 0.042669 -0.102709 -0.103231 -0.144040 0.040766 0.062360
D_AD_ORIT -0.049489 0.555501 0.555960 0.861266 1.000000 0.011964 0.020672 -0.059277 -0.075661 -0.155752 0.011588 0.052345
K_BLOOD -0.002999 -0.004314 -0.011671 0.030007 0.011964 1.000000 0.300430 0.023802 0.051519 0.012584 0.009055 -0.024080
NA_BLOOD 0.031139 -0.008899 0.012838 0.042669 0.020672 0.300430 1.000000 -0.005731 -0.022231 0.015816 -0.013838 0.012107
ALT_BLOOD -0.104688 -0.045470 -0.056683 -0.102709 -0.059277 0.023802 -0.005731 1.000000 0.519449 0.044393 -0.007868 0.053766
AST_BLOOD -0.053533 -0.083252 -0.057641 -0.103231 -0.075661 0.051519 -0.022231 0.519449 1.000000 0.077660 -0.030673 0.032433
L_BLOOD 0.003120 -0.172723 -0.125900 -0.144040 -0.155752 0.012584 0.015816 0.044393 0.077660 1.000000 0.005169 -0.004620
ROE 0.214393 0.005249 -0.022010 0.040766 0.011588 0.009055 -0.013838 -0.007868 -0.030673 0.005169 1.000000 0.041057
target 0.146107 0.072490 0.059376 0.062360 0.052345 -0.024080 0.012107 0.053766 0.032433 -0.004620 0.041057 1.000000 
plt.figure(figsize=(8,6))

sns.heatmap(
    corr_matrix,
    annot=True,
    cmap="coolwarm",
    fmt=".2f"
)

plt.title("Correlation Matrix")
plt.show()
../_images/58971cdf48d4714e7dae439c05c2d8208ec9004eaae765bd1bda0e88cb43f73e.png

Step 3: Feature preprocessing#

Before training machine learning models we need to define which variables should be used as predictors.

The dataset contains more than 100 potential input variables.

However, not all variables are suitable for the prediction of .

We apply three feature selection steps:

  1. Remove identifiers and target variables

  2. Remove variables with excessive missing values

  3. Inspect correlations between continuous variables

Important note:

Tree-based models such as Random Forest and XGBoost are generally robust to correlated predictors, therefore we usually do not automatically remove correlated features unless they are redundant.

Handling Missing Data#

Clinical datasets often contain missing values because:

  • laboratory tests may not be performed

  • documentation may be incomplete

  • measurements may not be available at admission

Therefore, we apply different strategies depending on the variable type:

  • All variables: Removed if more than 30% of values are missing

missing_percent = df.isna().mean()

# Keep continuous variables with <= 30% missing
continuous_keep = [
    col for col in continuous_features
    if missing_percent[col] <= 0.3
]

# Keep categorical/binary only if no missing values
categorical_keep = [
    col for col in categorical_features
    if missing_percent[col] <= 0.3
]

binary_keep = [
    col for col in binary_features
    if missing_percent[col] <= 0.3
]

# Combine selected features
selected_features = continuous_keep + categorical_keep + binary_keep + ["target"]
df = df[selected_features]

print("After removing binary and categorical features with missing values and continuous features with more than 30% missing values, there are {} features left.".format(len(selected_features) - 1))

After removing binary and categorical features with missing values and continuous features with more than 30% missing values, there are 95 features left.

# Split features and target
X = df.drop(columns="target")
y = df["target"]

Train-Test Split#

To evaluate the model properly we split the dataset into:

  • training data – used to train the models

  • test data – used to evaluate performance

We use stratified sampling to preserve the class distribution.

X_train, X_test, y_train, y_test = train_test_split(
    X, y,
    test_size=0.2,
    stratify=y,
    random_state=42
)

Preprocessing pipeline#

  • Data imputation

  • Standardization

# Continuous features
continuous_pipeline = Pipeline([
    ("imputer", SimpleImputer(strategy="median")),   # fill missing values
    ("scaler", StandardScaler())                     # normalize values
])

# Categorical + Binary features
categorical_pipeline = Pipeline([
    ("imputer", SimpleImputer(strategy="most_frequent"))  # fill missing values
])

# Combine both pipelines
preprocessor = ColumnTransformer([
    ("num", continuous_pipeline, continuous_keep),
    ("cat", categorical_pipeline, categorical_keep + binary_keep)
])

Encoding of categorical features is not necessary for this dataset, because all features are already encoded.

Step 4: Modeling#

We compare three different machine learning algorithms:

Logistic Regression

  • simple baseline model

  • interpretable

Random Forest

  • ensemble of decision trees

  • captures non-linear relationships

XGBoost

  • gradient boosting algorithm

  • often performs very well on tabular data

logreg_pipeline = Pipeline([
    ("preprocessing", preprocessor),
    ("model", LogisticRegression(max_iter=2000, class_weight="balanced"))
])

rf_pipeline = Pipeline([
    ("preprocessing", preprocessor),
    ("model", RandomForestClassifier(class_weight="balanced"))
])

xgb_pipeline = Pipeline([
    ("preprocessing", preprocessor),
    ("model", XGBClassifier(
        eval_metric="logloss",
        scale_pos_weight=scale_pos_weight
    ))
])

Hyperparameter Tuning#

Machine learning models have parameters that must be chosen before training.

logreg_grid = {
"model__C":[0.01,0.1,1,10]
}

rf_grid = {

"model__n_estimators":[100,300],
"model__max_depth":[5,10,None],
"model__min_samples_split":[2,5]

}

xgb_grid = {

"model__n_estimators":[100,300],
"model__max_depth":[3,6],
"model__learning_rate":[0.01,0.1]

}

print("Columns in X:")
print(X.columns.tolist())

Columns in X:
['AGE', 'S_AD_ORIT', 'D_AD_ORIT', 'K_BLOOD', 'NA_BLOOD', 'ALT_BLOOD', 'AST_BLOOD', 'L_BLOOD', 'ROE', 'INF_ANAM', 'STENOK_AN', 'FK_STENOK', 'IBS_POST', 'GB', 'DLIT_AG', 'ZSN_A', 'ant_im', 'lat_im', 'inf_im', 'post_im', 'TIME_B_S', 'SEX', 'SIM_GIPERT', 'nr_11', 'nr_01', 'nr_02', 'nr_03', 'nr_04', 'nr_07', 'nr_08', 'np_01', 'np_04', 'np_05', 'np_07', 'np_08', 'np_09', 'np_10', 'endocr_01', 'endocr_02', 'endocr_03', 'zab_leg_01', 'zab_leg_02', 'zab_leg_03', 'zab_leg_04', 'zab_leg_06', 'O_L_POST', 'K_SH_POST', 'MP_TP_POST', 'SVT_POST', 'GT_POST', 'FIB_G_POST', 'IM_PG_P', 'ritm_ecg_p_01', 'ritm_ecg_p_02', 'ritm_ecg_p_04', 'ritm_ecg_p_06', 'ritm_ecg_p_07', 'ritm_ecg_p_08', 'n_r_ecg_p_01', 'n_r_ecg_p_02', 'n_r_ecg_p_03', 'n_r_ecg_p_04', 'n_r_ecg_p_05', 'n_r_ecg_p_06', 'n_r_ecg_p_08', 'n_r_ecg_p_09', 'n_r_ecg_p_10', 'n_p_ecg_p_01', 'n_p_ecg_p_03', 'n_p_ecg_p_04', 'n_p_ecg_p_05', 'n_p_ecg_p_06', 'n_p_ecg_p_07', 'n_p_ecg_p_08', 'n_p_ecg_p_09', 'n_p_ecg_p_10', 'n_p_ecg_p_11', 'n_p_ecg_p_12', 'fibr_ter_01', 'fibr_ter_02', 'fibr_ter_03', 'fibr_ter_05', 'fibr_ter_06', 'fibr_ter_07', 'fibr_ter_08', 'GIPO_K', 'GIPER_NA', 'NITR_S', 'LID_S_n', 'B_BLOK_S_n', 'ANT_CA_S_n', 'GEPAR_S_n', 'ASP_S_n', 'TIKL_S_n', 'TRENT_S_n']

print("Continuous features:")
print(continuous_features)

Continuous features:
['AGE', 'S_AD_KBRIG', 'D_AD_KBRIG', 'S_AD_ORIT', 'D_AD_ORIT', 'K_BLOOD', 'NA_BLOOD', 'ALT_BLOOD', 'AST_BLOOD', 'L_BLOOD', 'ROE']

Step 5: Model Evaluation#

models = {
    "Logistic Regression": logreg_search.best_estimator_,
    "Random Forest": rf_search.best_estimator_,
    "XGBoost": xgb_search.best_estimator_
}

results = []

for name, model in models.items():

    y_prob = model.predict_proba(X_test)[:,1]

    fpr, tpr, thresholds = roc_curve(y_test, y_prob)
    optimal_idx = np.argmax(tpr - fpr)
    optimal_threshold = thresholds[optimal_idx]

    y_pred_optimal = (y_prob >= optimal_threshold).astype(int)

    report = classification_report(y_test, y_pred_optimal, output_dict=True)

    results.append({
        "Model": name,
        "ROC_AUC": roc_auc_score(y_test, y_prob),
        "Optimal_Threshold": optimal_threshold,
        "Precision": report["1"]["precision"],
        "Recall": report["1"]["recall"],
        "F1_score": report["1"]["f1-score"]
    })

results_df = pd.DataFrame(results).round(3)
results_df
Model ROC_AUC Optimal_Threshold Precision Recall F1_score
0 Logistic Regression 0.660 0.452 0.344 0.696 0.460
1 Random Forest 0.711 0.415 0.320 0.911 0.474
2 XGBoost 0.737 0.440 0.361 0.709 0.479

Confusion Matrix#

The confusion matrix shows the number of:

  • True Positives

  • True Negatives

  • False Positives

  • False Negatives

In clinical applications, false negatives may be particularly critical, because a high-risk patient might be missed.

fig, axes = plt.subplots(1, 3, figsize=(15,4))

for ax, (name, model) in zip(axes, models.items()):

    y_prob = model.predict_proba(X_test)[:,1]

    fpr, tpr, thresholds = roc_curve(y_test, y_prob)
    optimal_idx = np.argmax(tpr - fpr)
    optimal_threshold = thresholds[optimal_idx]

    y_pred = (y_prob >= optimal_threshold).astype(int)

    cm = confusion_matrix(y_test, y_pred)

    sns.heatmap(
        cm,
        annot=True,
        fmt="d",
        cmap="Blues",
        xticklabels=["No CHF","CHF"],
        yticklabels=["No CHF","CHF"],
        ax=ax
    )

    ax.set_title(name)
    ax.set_xlabel("Predicted")
    ax.set_ylabel("Actual")

plt.tight_layout()
plt.show()
../_images/5881cfd143e1a0ca32bb4129809ccce6361769fa94a4479e9766b4ce678291be.png 
fig, ax = plt.subplots(figsize=(7,6))

for name, model in models.items():

    RocCurveDisplay.from_estimator(
        model,
        X_test,
        y_test,
        ax=ax,
        name=name
    )

# random classifier baseline
ax.plot([0,1], [0,1], linestyle="--", color="grey")

ax.set_title("ROC Curve Comparison")
ax.set_xlabel("False Positive Rate")
ax.set_ylabel("True Positive Rate")

plt.legend()
plt.show()
../_images/4f0a1906fb277130e11a7f8f1dc2020fefce40d6e95241685a442bbee9093a73.png

Optional: Model Explainability#

Machine learning models are often considered “black boxes”. To build trust in clinical AI, it is important to understand how models make predictions.

We use two approaches:

  • Feature Importance – which variables are most influential overall

  • SHAP Values – how individual features influence predictions

SHAP values allow us to interpret predictions at both the global (model-level) and local (patient-level) scale.

rf_pipeline = rf_search.best_estimator_

rf_model = rf_pipeline.named_steps["model"]

feature_names = rf_pipeline.named_steps["preprocessing"].get_feature_names_out()

importances = rf_model.feature_importances_

feature_importance = pd.DataFrame({
    "feature": feature_names,
    "importance": importances
}).sort_values("importance", ascending=False)

feature_importance.head(20)
feature importance
15 cat__ZSN_A 0.157926
0 num__AGE 0.092992
8 num__ROE 0.052270
3 num__K_BLOOD 0.041858
37 cat__endocr_01 0.040864
6 num__AST_BLOOD 0.038621
17 cat__lat_im 0.037699
4 num__NA_BLOOD 0.036896
7 num__L_BLOOD 0.033988
40 cat__zab_leg_01 0.032260
1 num__S_AD_ORIT 0.028063
2 num__D_AD_ORIT 0.023794
5 num__ALT_BLOOD 0.021469
20 cat__TIME_B_S 0.021380
47 cat__MP_TP_POST 0.019474
21 cat__SEX 0.018108
13 cat__GB 0.015928
16 cat__ant_im 0.015712
10 cat__STENOK_AN 0.015692
90 cat__ANT_CA_S_n 0.015347 
top_features = feature_importance.head(20)

plt.figure(figsize=(8,6))

sns.barplot(
    data=top_features,
    x="importance",
    y="feature"
)

plt.title("Top 20 Feature Importances (Random Forest)")
plt.xlabel("Importance")
plt.ylabel("Feature")

plt.show()
../_images/612a15305b275ef95999594f883c7abe2cc11703519c20d8851ef62024eceaba.png

These features contribute most strongly to the model’s predictions.

Important note:

Feature importance does not necessarily indicate causality. Highly ranked variables simply help the model distinguish between patients with and without the outcome.

Discussion#

Questions to consider:

  • Which model performed best?

  • Which features were most important?

  • Are these findings clinically plausible?

  • What additional data could improve predictions?