Work sharing: My Stroke Predict Analyses & Models
Github: https://github.com/pkodja/StrokePredict/tree/main/README.md
Coding environment: CONDA VIRTUAL ENVIRONMENT
(Confidential Source) - Use only for educational purposes
Size: contains 5110 observations (rows) and 11 clinical features for predicting stroke events
File type: healthcare-dataset-stroke-data.csv
Author: fedesoriano
License: Data files © Original Authors
According to the World Health Organization (WHO) stroke is the 2nd leading cause of death globally, responsible for approximately 11% of total deaths. This dataset is used to predict whether a patient is likely to get stroke based on the input parameters like gender, age, various diseases, and smoking status. Each row in the data provides relevant information about the patient.
- id: unique identifier
- gender: "Male", "Female" or "Other"
- age: age of the patient
- hypertension: 0 if the patient doesn't have hypertension, 1 if the patient has hypertension
- heart_disease: 0 if the patient doesn't have any heart diseases, 1 if the patient has a heart disease
- ever_married: "No" or "Yes"
- work_type: "children", "Govt_jov", "Never_worked", "Private" or "Self-employed"
- Residence_type: "Rural" or "Urban"
- avg_glucose_level: average glucose level in blood
- bmi: body mass index
- smoking_status: "formerly smoked", "never smoked", "smokes" or "Unknown"*
- stroke: 1 if the patient had a stroke or 0 if not
*Note: "Unknown" in smoking_status means that the information is unavailable for this patient
The objective of this academic project is to use data scientist techniques to preprocess and analyze the data structure to discover relevant features to use as inputs for the machine learning process to ultimately select an adequate model for the predictions of stroke events.
In order to achieve the defined objective, this study will go through four main processes before concluding:
- Data exploration and analysis
- Feature Engineering and Feature Selection
- Model Training
- Model Evaluation
The column stroke is the target feature or variable and the rest of columns are explanatory variables or features.
The structure of the dataset will be analyzed to prepare clean and complete data for the following feature analysis.
As said above, there are 12 features with one target feature or response variable -stroke- and 11 explanatory variables. The dataset have:
- 4 numerical variables: "id", "age", "avg_glucose_leve" and "bmi"
- 8 categorical variables with 3 ordinal variables and 5 nominal variables:
- Ordinal variables: "hypertension", "heart_disease" and "stroke" (Target Feature)
- Nominal variables: "gender", "ever_married", "work_type", "Residence_type" and "smoking_status"
By observing the dataset "stroke" column structure, it looks strongly unbalanced:
stroke
0 4861
1 249
Name: count, dtype: int64
There are more unstroked people than stroked ones. The 5110 observations are not much enough to have a balanced dataset
The column "id" will be used as the dataset row indexes and not as a pure explanatory variable. It remains then 10 variables for data analysis.
The dataset has a single row with "gender" column label "Other". This line will be deleted as it could not have a significant impact on the entire dataset. The dataset size will then be 5109
The column "bmi" has missing data for 201 observations:
df_PrePro.isna().sum()
gender 0
age 0
hypertension 0
heart_disease 0
ever_married 0
work_type 0
Residence_type 0
avg_glucose_level 0
bmi 201
smoking_status 0
stroke 0
dtype: int64
Among the 201 observations, there are 40 strokes, so these observations cannot be removed because of the small number of people who suffered a stroke in the entire dataset. Their ""bmi" will be imputed with the dataset stroked observations bmi mean and the rest 161 will have their bmi imputed with the dataset unstroked people bmi mean.
All the categorical features are unbalanced data exactly like for "stroke" above. There more unstroked people than stroked per lable and will lead model to predict within unstroked observations:
Ordinal feature example:
Nominal feature example:
Apart from the feature "age" which is a balanced one all the remaining numerical features are imbalance. Here are some of their characteristics:
-
Age:
-
avg_glucose_level: have outliers and a second mode in outliers above the maximum limit of glucose level average 125. This could be due to the small size of the dataset and will have negative impact on predictions. The model is likely to predict within unstroked observations
-
bmi: Have many outliers too This could be due to the small size of the dataset and will have negative impact on predictions. Here too, he model is likely to predict within unstroked observations
-
Hypertension, avg_glucose_level, ever_married and stroke Married people with high average glucose level than normal have high stroke rate. The risk seems less for the same people with hypertension
-
Hypertension, avg_glucose_level, heart_disease and stroke
As we discussed earlier, people with average glucose levels between 70 and 125 have a high rate of stroke,
regardless of heart disease or hypertension. Within them, those with hypertension also a bit more chance to face stroke. -
Hypertension, age, avg_glucose_level and stroke
Most people seem to have an average blood sugar below 125, regardless of age, but strokes occur among them in people over 50 years old.
The data imbalance does not allow to really notice the hypertension impact on those people but we have the same trend for old people with hypertension.
The nominal categorical features were normalized in ordinal to allow correlations study.
Pandas function corr() uses Pearson process which requires numerical variables.
As nominal categorical variables were normalized in ordinal pd.corr() is used to an overview on the relationship between all the features.
Then adequate statistic tests were used to qualify features relationship as it follows:
-
Correlation between numerical variable with Pearson statistic test with pd.corr()
-
Correlation between categorical variables with Chi2 statistic test with chi2_contingency() function:
chi2CorrTest(df_transf["hypertension"],df_transf["stroke"]) Chi2 Test P_value: 1.688936253410575e-19 1 These features are ' correlated' as ''P_value'' < 0.05 -
Correlation between numerical and categorical variables with ANOVA test through "SelectKBest" class:
#import statsmodels.api resulta=statsmodels.formula.api.ols('age~work_type',data=df_transf).fit() corrAgeWorkType=statsmodels.api.stats.anova_lm(resulta) corrAgeWorkType df sum_sq mean_sq F PR(>F) work_type 4.0 1.215328e+06 303831.964265 1110.246464 0.0 Residual 5104.0 1.396769e+06 273.661727 NaN NaN #import statsmodels.api resulta=statsmodels.formula.api.ols('age~ever_married',data=df_transf).fit() corrAgeWorkType=statsmodels.api.stats.anova_lm(resulta) corrAgeWorkType df sum_sq mean_sq F PR(>F) ever_married 1.0 1.204583e+06 1.204583e+06 4370.69022 0.0 Residual 5107.0 1.407514e+06 2.756048e+02 NaN NaN
Feature Engineering conclusion: above P-values (0.0 < 0.05) confirm correlations between "age" and "ever_married" and "work_type"; but as pd.corr() has previously proved the correlation coefficient of "age" and "ever_married" is high, implying a strong relationship between both. So to avoid data redundancy a choice might be made between "age" and "ever_married".
Finally we noticed that only 4 explanatory features -age, hypertension, heart_disease, avg_glucose_level- have significant correlation with the target feature "stroke".
For this study 2 models will be tested fondamentally based on the use of pd.corr() and SelectKBest():
- Model1: Stroke~age,hypertension,heart_disease,avg_glucose_level (By prefering age and leaving out ever_married strongly correlated)
- Model2 Stroke~age,hypertension,heart_disease,avg_glucose_level,ever_married (adding ever_married to Model1)
The best will be kept.
Model1: Stroke ~ age,hypertension,heart_disease,avg_glucose_level Model2 Stroke ~ age,hypertension,heart_disease,avg_glucose_level,ever_married
Python classifier models LogisticRegression, MLPClassifier, DecisionTreeClassifier and RandomForestClassifier were used for the data training and prediction.
For the process, the stroke dataset was splitted in training and testing datasets in 80/20 rate. Each part has its target feature -stroke- and explanatory features. The testing dataset is used for the prediction and the mode evaluation.
The main process was first managed with the LogisticRegression Classifier model to refine the training and testing datasets. Logistic regression class_weight='balanced' was used at first due to high data imbalance, then SMOTE was used. to generate observations for the training and testing datasets before the new datasets are used on subsequent classifiers
Check the Logistic Regression model results below:
Model: LogisticRegression(max_iter=1000, random_state=0)
No class_weight='balanced' use
------------------------------
cf.loc[:,:] = confusion_matrix(y_true= y_test,y_pred= y_test_predicted)
cf
Results without class_weight='balanced' and SMOTE use:
------------------------------------------------------
y_test_0 y_test_1
y_pred_0 971 0
y_pred_1 51 0
-----------------------------------------------------------
Model: LogisticRegression(class_weight='balanced', max_iter=1000, random_state=0)
results with class_weight='balanced' and before SMOTE use:
---------------------------------------------------------
y_test_0 y_test_1
y_pred_0 748 223
y_pred_1 14 37
results after SMOTE use without class_weight='balanced':
------------------------------------------------------
y_test_0 y_test_1
y_pred_0 758 213
y_pred_1 257 714
After model training performance, the Model1: Stroke ~ age,hypertension,heart_disease,avg_glucose_level is selected.
Among the selected features there is "avg_glucose_level" feature with outliers. Those outliers -observations- was removed and their positive effects are shown below in the model evaluation.
- avg_glucose_level data status before outliers cancelling:
- avg_glucose_level data status after outliers cancelling:
**With outliers**
precision recall f1-score support
0 0.75 0.78 0.76 971
1 0.77 0.74 0.75 971
accuracy 0.76 1942
macro avg 0.76 0.76 0.76 1942
weighted avg 0.76 0.76 0.76 1942
-----------------------------------------------------
**Without outliers**
precision recall f1-score support
0 0.75 0.76 0.76 869
1 0.76 0.75 0.75 869
accuracy 0.75 1738
macro avg 0.75 0.75 0.75 1738
weighted avg 0.75 0.75 0.75 1738
The macro avg values have increased a bit. Recall for stroked has increase a bit too and so is for f1 score macro avg
**With outliers**
Report of Decision Tree classifier model
precision recall f1-score support
0 0.69 0.93 0.79 971
1 0.90 0.58 0.71 971
accuracy 0.76 1942
macro avg 0.79 0.76 0.75 1942
weighted avg 0.79 0.76 0.75 1942
------------------------------------------------------
**Without outliers**
precision recall f1-score support
0 0.70 0.95 0.81 869
1 0.92 0.60 0.73 869
accuracy 0.78 1738
macro avg 0.81 0.78 0.77 1738
weighted avg 0.81 0.78 0.77 1738
Conclusion for Decisional Tree: Better improvement after outliers cancelling of f1-score, Recall and Precision
**With outliers**
Report of Random Forest model
precision recall f1-score support
0 0.71 0.96 0.82 971
1 0.93 0.61 0.74 971
accuracy 0.78 1942
macro avg 0.82 0.78 0.78 1942
weighted avg 0.82 0.78 0.78 1942
------------------------------------------------------
**Without outliers**
precision recall f1-score support
0 0.72 0.95 0.82 869
1 0.93 0.63 0.75 869
accuracy 0.79 1738
macro avg 0.82 0.79 0.79 1738
weighted avg 0.82 0.79 0.79 1738
Conclusion for Random Forest: seems not well performing. Not better improvement after outliers cancelling then
In general the Decisional Tree Model seems the best after data adjustment with SMOTE and average glucose level outliers cancelling. The Logistic Regression model has also improved a little bit proving that simple classification model can handle observations classification correctly.
Before the outliers treatment, both Decisional Tree and Random Forest were the best.
The small size of the dataset, data imbalance and outliers have definitely negative impact on the model training.
The decisional tree final model will be saved on disk for further use of its current parameters