These are running environments and may work with lower specs.
OS : Windows 10
CPU : Intel Core i7-8700K @ 3.70GHz
RAM : 16GB
GPU : NVIDIA Geforce GTX 1660
The environment was mainly built using anaconda.
conda : 4.9.2
python : 3.8.5
Install anaconda from the following URL.
https://www.anaconda.com/products/individual
Run the following command at the command prompt to install the necessary python libraries.
$ pip install -r requirements.txt
The installation time is about 10 minutes.
The following can be executed in "demo.ipynb".
The run time is about 30 seconds.
Get training data and testing data.
The data format can be "pandas.dataframe" or "numpy.array".
import pandas as pd
train_data = pd.read_csv(r'.\data\traindata_overtANDlatentThyroidism.csv', encoding='cp932')
test_data = pd.read_csv(r'.\data\testdata_overtThyroidism.csv', encoding='cp932')Extract the target columns from the data.
Also, extract or exclude specified attributes from the data as needed.
num_features = ['AST', 'ALT', 'γ-GTP', 'Total_cholesterol', 'RBC', 'Hb', 'UA', 'S-Cr', 'UA_S-Cr', 'ALP']
cat_features = ['Sex']
obj_variable = 'class'
info_variable = 'attribute'
target_columns = num_features+cat_features+[obj_variable, info_variable]
#Extract "target_columns"
train_data = train_data.loc[:,target_columns]
test_data = test_data.loc[:,target_columns]
#Exclude "info_variable" including "gunma" from the training data
train_data = train_data[~train_data[info_variable].str.contains('gunma')].reset_index(drop=True)"num_features" : Numerical features to be input to the machine learning model.
"cat_features" : Categorical features to be input to the machine learning model.
"obj_variable" : Teacher labels for machine learning models.
"info_variable" : Attribute information for data extraction.
Apply label encoding to categorical features and teacher labels.
#Label encoding("cat_features")
label_encoder = {'male':0, 'female':1}
train_data.loc[:,cat_features] = train_data.loc[:,cat_features].applymap(lambda x: label_encoder[x])
test_data.loc[:,cat_features] = test_data.loc[:,cat_features].applymap(lambda x: label_encoder[x])
#Label encoding("obj_variable")
label_encoder = {'hyper':1, 'hypox':0, 'normal':0}
train_data[obj_variable] = train_data[obj_variable].map(lambda x: label_encoder[x])
test_data[obj_variable] = test_data[obj_variable].map(lambda x: label_encoder[x])Define model.
Here is an example with "CatBoost".
from catboost import CatBoostClassifier
##Define machine learning model
model = CatBoostClassifier() #CatboostThe training data and test data are divided into equal numbers, and training and testing are performed on each divided data.
Here is an example where the data is divided into 10 parts.
import numpy as np
from sklearn.model_selection import StratifiedKFold
from sklearn.metrics import confusion_matrix, roc_auc_score
kf = StratifiedKFold(n_splits=10)
X_train = train_data.loc[:,num_features+cat_features]
y_train = train_data[obj_variable]
X_test = test_data.loc[:,num_features+cat_features]
y_true = test_data[obj_variable]
result = pd.DataFrame()
for train_indexes, test_indexes in zip(kf.split(X_train, y_train), kf.split(X_test, y_true)):
train_index = train_indexes[0]
test_index = test_indexes[1]
model.fit(X_train.loc[train_index,:], y_train[train_index], verbose=0)
proba = model.predict_proba(X_test.loc[test_index,:])
positive_proba = proba[:,1]
auroc = roc_auc_score(y_true[test_index], positive_proba)
y_pred = np.where(positive_proba>=0.5, 1, 0)
cm = confusion_matrix(y_true[test_index], y_pred, labels=[0, 1])
tn, fp, fn, tp = cm.flatten()
recall = tp / (tp+fn)
specificity = tn / (tn+fp)
result = result.append(pd.Series([auroc, recall, specificity], index=['AUROC', 'Recall', 'Specificity']), ignore_index=True)- Our article
https://www.medrxiv.org/content/10.1101/2021.03.30.21254605v1 - Web sites about software
https://www.anaconda.com/products/individual - Web sites about machine learning related python libraries
https://scikit-learn.org/stable/supervised_learning.html#supervised-learning
https://www.tensorflow.org/api_docs/python/tf/keras?hl=ja