Cardiotocography (CTG) is a method commonly used to assess fetal heart rate with ultrasound during labour, and is often used to determine if a fetus is at risk of hypoxia. We evaluated methods on two different datasets to create models that accurately predict the health of a fetus. Our models on the first dataset include testing several classification models to predict/classify the health of the fetus, and identify which model is most effective. On the second dataset, we preprocess biosignals into usable data, then represent as images and train a DCNN to predict fetal health. We also venture into feature extraction to test the efficacy of methods used in the first dataset on the second
| Tabulated Dataset | CTU-CHB dataset | Slack channel | Project report | Timesheet |
https://drive.google.com/file/d/1HbFV1k-9HB5zfbAsMjozX7fhjIk0FmpX/view
fetal-health-classification
├── CTU-CHB ## CTU-CHB database
│ ├── dat ## Unprocessed raw dat waveforms
│ ├── dat_cwt ## cwt plots
│ ├── dat_recurrence_plots ## recurrence plots
│ ├── dat_spectrogram ## spectrogram plots
│ ├── DCNN ## DCNN Model training and testing
│ ├── Feature Extraction ## Feature extraction files for tabulation
│ ├── generate_plots ## Files for generating plots
│ ├── processed_dat ## Processed and cleaned waveforms
│ ├── outcomes.csv ## CTU-CHB metadata and labels
│ ├── waveform_processing.py ## processing waveforms
|
├── TabulatedCTG ## Tabulated CTG dataset
│ ├── Classify CTU-CHB ## ML methods on CTU-CHB
│ ├── fetal_health.csv ## Tabulated Dataset
│ ├── ML Methods.py ## Machine learning methods
│ ├── NN.py ## Neural Network classifier
|
├── requirements.txt ## Setup requirements
Ensure requirements in the Installation section are met.
Comparing different machine learning models with various physiology signals to assess the status of fetuses.
To run ML methods on fetal_health.csv: python TabulatedCTG/ML_Methods.py
Analysis of FHR waveforms with the Intrapartum Cardiotocography Database.
- Run
python3 CTU-CHB/waveform_processing.pyon dataset downloaded into/dat
Example signal before and after preprocessing:
- Run the program:
python3 CTU-CHB/generate_plots/generate_recurrence_plots.py
Here is one sample recurrence plot:
To just test the DCNN model, run the test on the prepared test split: python CTU/CHB/DCNN/python test.py
Instructions for training the model are in the Reproduction section below.
- Feature extraction:
python3 "CTU-CHB/Feature Extraction/Feature_Extraction.py" - Test ML methods from tabulated dataset on new dataset:
python3 "TabulatedCTG/Classify_CTU-CHB.py"
Install requirements: pip install -r requirements.txt
- DCNN requires tensorflow to be setup.
Download the fetal-health.csv: https://www.kaggle.com/datasets/andrewmvd/fetal-health-classification/data
- Place into
TabulatedCTG/.
Download the CTU-CHB data: wget -r -N -c -np https://physionet.org/files/ctu-uhb-ctgdb/1.0.0/
- Place waveform .dat files into
dat/.
Comparing different machine learning models with various physiology signals to assess the status of fetuses.
To run ML methods on fetal_health.csv: python TabulatedCTG/ML_Methods.py
# Select columns for ML methods
df = df[['baseline value', 'accelerations', 'fetal_movement', 'uterine_contractions', 'light_decelerations',
'severe_decelerations', 'prolongued_decelerations', 'abnormal_short_term_variability',
'mean_value_of_short_term_variability', 'percentage_of_time_with_abnormal_long_term_variability',
'mean_value_of_long_term_variability', 'fetal_health']]
# Create X and Y data
X = df.iloc[:, :-1]
y = df.iloc[:, -1]
# Split the data into 75% training and 25% testing data
X_train, X_test, y_train, y_test = train_test_split(X, y)
model.fit(X_train, y_train)
y_pred = model.predict(X_test)
# Output the accuracies
print(model)
print(classification_report(y_test, y_pred))
print("Balanced Accuracy: ", balanced_accuracy_score(y_test, y_pred))
# Output the confusion matrix
cm = confusion_matrix(y_test, y_pred, labels=[1,2,3])
ConfusionMatrixDisplay(cm).plot()- DecisionTree:
precision recall f1-score support
1.0 0.95 0.95 0.95 418
2.0 0.72 0.78 0.75 67
3.0 0.90 0.81 0.85 47
accuracy 0.92 532
macro avg 0.86 0.85 0.85 532
weighted avg 0.92 0.92 0.92 532
Balanced Accuracy: 0.8463918319428005
- GaussianNB:
GaussianNB()
precision recall f1-score support
1.0 0.97 0.86 0.91 416
2.0 0.49 0.86 0.63 79
3.0 0.69 0.49 0.57 37
accuracy 0.83 532
macro avg 0.72 0.74 0.70 532
weighted avg 0.88 0.83 0.84 532
Balanced Accuracy: 0.7351396856934831
- RandomForest:
RandomForestClassifier(n_estimators=200)
precision recall f1-score support
1.0 0.95 0.99 0.97 416
2.0 0.89 0.72 0.79 78
3.0 0.95 0.97 0.96 38
accuracy 0.95 532
macro avg 0.93 0.89 0.91 532
weighted avg 0.94 0.95 0.94 532
Balanced Accuracy: 0.8924032838506523
- KNeighbors Model:
precision recall f1-score support
1.0 0.91 0.96 0.94 415
2.0 0.78 0.61 0.68 69
3.0 0.85 0.71 0.77 48
accuracy 0.89 532
macro avg 0.85 0.76 0.80 532
weighted avg 0.89 0.89 0.89 532
Balanced Accuracy: 0.7602948023979978
- Feature importance: Using Random Forest model to calculate the feature importance:
features = ['baseline value','accelerations','fetal_movement','uterine_contractions','light_decelerations','severe_decelerations','prolongued_decelerations','abnormal_short_term_variability','mean_value_of_short_term_variability','percentage_of_time_with_abnormal_long_term_variability','mean_value_of_long_term_variability','histogram_width','histogram_min','histogram_max','histogram_number_of_peaks','histogram_number_of_zeroes','histogram_mode','histogram_mean','histogram_median','histogram_variance','histogram_tendency']
f_i = list(zip(features,rf.feature_importances_))
f_i.sort(key = lambda x : x[1])
plt.barh([x[0] for x in f_i],[x[1] for x in f_i])
plt.show()
The result clearly shows that proloungued deceleration and uterine contractions are important features to be considered about
Analysis of fetal statuses with the Intrapartum Cardiotocography Database
- Download dataset:
wget -r -N -c -np https://physionet.org/files/ctu-uhb-ctgdb/1.0.0/ - Move dataset:
mv physionet.org/files/ctu-uhb-ctgdb/1.0.0/* CTU-CHB/dat/ - Run
python3 waveform_processing.pyon dataset downloaded into/dat
Get the FHR signals by processing the CTU-CHB data.
Convert the signals to waveform:
# the total seconds of the file is the length of the file divided by 4 as it was sampled at 4Hz
ts = np.arange(len(fetal_hr))/4.0
#get the valid segments (ie the processed segments without the long gaps)
selected_segments = get_valid_segments(fetal_hr, ts, FILEPATH, verbose=False,
#max_change=15, verbose_details=True
)
len_s = len(selected_segments)
new_signal = []
#add the segments together for the new processed signal
for i in range(len_s):
new_signal.extend(selected_segments[i]['seg_hr'])
- Recurrence Plots:
python3 CTU-CHB/generate_plots/generate_recurrence_plots.py
The recurrence plot function used here:
# Based on: https://stackoverflow.com/questions/33650371/recurrence-plot-in-python
def rec_plot(data, eps=0.3, steps=15):
d = pdist(data[:, None])
d = np.floor(d / eps)
d[d > steps] = steps
Z = squareform(d)
return ZSpectrograms and CWT:
- In matlab, run
generate_spectrograms.morgenerate_cwts.m
Training: This assumes the data has been preprocessed, image plots have been generated, and you are in the CTU-CHB/DCNN directory:
python process_load_label.py- Process, load and labelpython generate.py <type>- Generate test/train split: (where type is eitherdat_spectrogram,dat_cwt, ordat_recurrence_plots)python train.py <num of epochs>- Train Model:, 10-50 epochs recommended
To load model (which is seperated into architecture and weights):
with open('model_architecture.json', 'r') as json_file:
architecture = json.load(json_file)
model = tf.keras.models.model_from_json(architecture)
model.load_weights('model.weights.h5')