/prediction_heart_failure

Model that predict the survive or not of people that had heart failure

Primary LanguageJupyter Notebook

Prediction Survive Heart Failure

Overview of project

In this project, was made a model of machine learning using the Logistic Regression for predicted the person that had heart failure would survive or not.

Project structure

  • data folder that contain the datas.
  • eda folder that contain the notebook of exploratory data analysis.
  • modeling folder that contain the notebook of modeling.
  • src folder that contain the images that are use in readme.

Description of problem

A hospital wants to develop a machine learning model to predict whether or not a person with heart failure will survive during the patient's follow-up period. To develop this model, the hospital has a database that contains information on patients who had heart failure.

Exploratory Data Analysis

About Dataset

This dataset contains the medical records of 299 patients who had heart failure, collected during their follow-up period, where each patient profile has 13 clinical features.

Link dataset: https://archive.ics.uci.edu/dataset/519/heart+failure+clinical+records

Insights

Kde_plot
Some observations:

  • The highest concentration of age is between 50 and 70 years.
  • The variable creatinine_phosphokinase is a distribution positive asymmetric. Because the normal values of creatinine_phosphokinase is 10 to 120 mcg/L, but there some people that are with the high level of creatinine_phosphokinase.
  • The variable serum_creatinine also is a distribution positive asymmetric. Because the normal values of serum_creatinine is 0.7 to 1.3 mg/dL, but there some people that are with the high level of serum_creatinine.
  • The variable platelets also is a distribution positive asymmetric a little softer. Because the normal values of platelets is 150 to 450 kiloplatelets/mL, but there few people that are with the high level of platelets.
  • The variable serum_sodium is a distribution negative asymmetric. Because the normal values of serum_sodium is 135 to 145 mEq/L, but there few people that are with the low level of serum_sodium.

Bar_plot
Some observations:

  • The numbers of men is almost double the number of women.
  • Most of the people that had heart failure, doesn't smoking.
  • Most of the people that had heart failure, doesn't had high blood pressure.

Heatmap_plot

  • The numerical variables has much little correlation each other.

Boxplot_plot
Some observations:

  • The mean of people that had heart failure and dead is 65 years, that show how older the person more chance of died . And also can be observed that only one person that had 90 years doesn't died.
  • There a high level of creatinine_phosphokinase doesn't indicate that the person that had heart failure will have more chance death. This also occur with serum_creatinine.
  • If the person had one mean of 160 days of follow-up or more, your survival chance is biggest.

Bar2_plot Some observations:

  • There high blood pressure didn't important fact for the person death that had heart failure, like also smoking, anaemia and diabets.
  • Of 194 men that had heart failure 68,04% suirvived and of 105 women that had heart failure 67,62% also survived. And this show that the sex of person didn't important fact for the person death.

Modeling

Was made a model with all features of dataset for predict response variable, after was made again a model applying feature selection for choice a better features for predict response variable.

Model with all features

metrics_allfeatures

cm_allfeatures Observing the confusion matrix above, the model with all features classified 4 false negative and this is bad, because the people that were classified as false negative dead.

precision_recall_allfeatures

  • AUC precision-recall curve : 88.25%
  • Can be observed that precision-recall curve and AUC show that precision is good for this model with all features.

Model with best features

metrics_best_features Comparing with a model with all features, the model with best features there a increased in the ROC AUC, but the Precision descreased this show that model with best features is classifying more instances like positives, but this is bad because the number of false positives also will increse.

cm_best_features Some observations:

  • As already mentioned above the model with best features is classifying more instances like positives and this also increase the number of false positive.
  • Both the models couldn't decrease the number of false negative, despite having classified just 4 false negative this is bad because show that 4 people dead due a wrong classification.

precision_recall_best_features

  • AUC precision-recall curve: 88.87%
  • Can be observed that precision-recall curve and AUC show that precision of model with best features is better for classify positives.

Conclusion

In this project, the goal was develop a model of machine learning for predict if a person that had heart failure during follow-up would survive or not. Using a model with a best features obtained a perfomance a little better than with all features. But the model with best features can't decrease the number of false positives, perhaps with a larger dataset and using technics more advanced like feature engineering the model would permormance better. Despite that, the model had a good performance with a dataset that used.