Online Retraining
Select a previously trained model.
Train a new predictive model for Sleep Staging or Sleep Disordered Breathing. While training is in progress, you may navigate away from the page or close your browser. Trained models may be saved for later use in the Analysis section of the platform.
Select an existing model to test on out-of-sample data, or configure and train a new model.
Sleep Disordered Breathing
Configuration
Select between training a binary classifier to predict whether the AHI of a child is at least 5 or to classify recording as Sleep or Wake, or a regressor to produce a point estimate of the AHI.
Select a Linear, Gaussian, or Polynomial kernel. The Linear kernel allows for better model interpretability, while the Gaussian kernel is more flexible and may capture non-linearities. Polynomial kernels offer a balance in complexity and the degree of non-linearity is configurable.Cost-sensitive Support Vector Classification may be used for imbalanced datasets. When different costs are applied, rescaling occurs so that the average C value across both classes is equal to the C value that is set during the hyperparameter search.
Model Information
Feature Selection
| Feature | Description | |
|---|---|---|
| Desats | Number of 3% drops per hour in the oxygen saturation. | |
| Desat Depth | Cumulative percentage drop per hour in the oxygen saturation during desaturations. | |
| Desat Duration | Cumulative duration per hour of desaturations to resaturation. | |
| Desat Area | Cumulative area per hour of desaturations to resaturation. | |
| OSA Band | Normalised spectral component of the oxygen saturation in the OSA Band (0.01-0.035Hz). | |
| OSA Band Mean | First statistical moment of the oxygen saturation power spectral density in the OSA Band (0.01-0.035Hz). | |
| OSA Band Variance | Second statistical moment of the oxygen saturation power spectral density in the OSA Band (0.01-0.035Hz). | |
| OSA Band Skewness | Third statistical moment of the oxygen saturation power spectral density in the OSA Band (0.01-0.035Hz). | |
| OSA Band Kurtosis | Fourth statistical moment of the oxygen saturation power spectral density in the OSA Band (0.01-0.035Hz). | |
| OSA Band Peak | Spectral component of the oxygen saturation in the OSA Band (0.01-0.035Hz) with the highest power. | |
| Sats Mean | First statistical moment of the oxygen saturation. | |
| Sats Variance | Second statistical moment of the oxygen saturation. | |
| Sats Skewness | Third statistical moment of the oxygen saturation. | |
| Sats Kurtosis | Fourth statistical moment of the oxygen saturation. | |
| Sats CTM1 | Central Tendency Measure of the oxygen saturation where the radius equals 1. A measure of variability. | |
| Sats CTM3 | Central Tendency Measure of the oxygen saturation where the radius equals 3. A measure of variability. | |
| Sats CTM6 | Central Tendency Measure of the oxygen saturation where the radius equals 6. A measure of variability. | |
| Sats Complexity | Lempel-Ziv Complexity of the binarised oxygen saturation. A measure of repeating sequences in the oxygen saturation. | |
| Sats RMSSD | Root mean square of successive differences in the oxygen saturation. RMSSD is commonly used as a measure of heart rate variability. | |
| Sats Shannon Entropy | Shannon entropy of the oxygen saturation. A static measure of regularity. | |
| Sats Permutation Entropy | Permutation entropy of the oxygen saturation where the embedding delay equals 1, and m equals 3. A dynamic measure of regularity. | |
| Sats Sample Entropy | Sample entropy of the oxygen saturation where the template length equals 2, and the tolerance equals the product of 0.2 and the standard deviation of the oxygen saturation. A dynamic measure of regularity. | |
| Sats Power | Total spectral component of the oxygen saturation. | |
| Sats Power Mean | First statistical moment of the oxygen saturation power spectral density. | |
| Sats Power Variance | Second statistical moment of the oxygen saturation power spectral density. | |
| Sats Power Skewness | Third statistical moment of the oxygen saturation power spectral density. | |
| Sats Power Kurtosis | Fourth statistical moment of the oxygen saturation power spectral density. | |
| Sats Power Peak | Spectral component of the oxygen saturation with the highest power. | |
| Sats DFA | Detrended fluctuation analysis of the oxygen saturation over 20 log-spaced windows, where the minimum window size is 10, and the maximum is 1000. A measurement of self-affinity. | |
| Sats AutoCorr 1 | First-order autocorrelation of the oxygen saturation. Measures the correlation of the oxygen saturation with a delayed copy of itself. | |
| Sats AutoCorr 2 | Second-order autocorrelation of the oxygen saturation. Measures the correlation of the oxygen saturation with a delayed copy of itself. | |
| Sats AutoCorr 3 | Third-order autocorrelation of the oxygen saturation. Measures the correlation of the oxygen saturation with a delayed copy of itself. | |
| Sats AutoCorr 4 | Fourth-order autocorrelation of the oxygen saturation. Measures the correlation of the oxygen saturation with a delayed copy of itself. | |
| Pulse Rate Rise Index | Continuous rises per hour in the pulse rate of at least 15 beats per minute, occuring over a period of at least 4 seconds. | |
| Pulse Rate VLF Band | Normalised spectral component of the pulse rate in the VLF Band (0.01-0.04Hz). | |
| Pulse Rate LF Band | Normalised spectral component of the pulse rate in the LF Band (0.04-0.15Hz). | |
| Pulse Rate HF Band | Normalised spectral component of the pulse rate in the HF Band (0.15Hz to 0.4Hz). The HF Band is attenuated toward the Nyquist frequency. | |
| Pulse Rate LF/HF | Ratio of spectral components in the LF (0.04-0.15Hz) and HF (0.15Hz to Nyquist) Bands of the pulse rate. An estimate of sympathovagal balance. | |
| Pulse Rate Mean | First statistical moment of the pulse rate. | |
| Pulse Rate Variance | Second statistical moment of the pulse rate. | |
| Pulse Rate Skewness | Third statistical moment of the pulse rate. | |
| Pulse Rate Kurtosis | Fourth statistical moment of the pulse rate. | |
| Pulse Rate CTM1 | Central Tendency Measure of the pulse rate where the radius equals 1. A measure of variability in the pulse rate. | |
| Pulse Rate CTM3 | Central Tendency Measure of the pulse rate where the radius equals 3. A measure of variability in the pulse rate. | |
| Pulse Rate CTM6 | Central Tendency Measure of the pulse rate where the radius equals 6. A measure of variability in the pulse rate. | |
| Pulse Rate Complexity | Lempel-Ziv Complexity of the binarised pulse rate. A measure of repeating sequences in the pulse rate. | |
| Pulse Rate RMSSD | Root mean square of successive differences in the pulse rate. RMSSD is commonly used as a measure of heart rate variability. | |
| Pulse Rate Shannon Entropy | Shannon entropy of the pulse rate. A static measure of regularity. | |
| Pulse Rate Permutation Entropy | Permutation entropy of the pulse rate where the embedding delay equals 1, and m equals 3. A dynamic measure of regularity. | |
| Pulse Rate Sample Entropy | Sample entropy of the pulse rate where the template length equals 2, and the tolerance equals the product of 0.2 and the standard deviation of the pulse rate. A dynamic measure of regularity. | |
| Pulse Rate Power | Total spectral component of the pulse rate. | |
| Pulse Rate Power Mean | First statistical moment of the pulse rate power spectral density. | |
| Pulse Rate Power Variance | Second statistical moment of the pulse rate power spectral density. | |
| Pulse Rate Power Skewness | Third statistical moment of the pulse rate power spectral density. | |
| Pulse Rate Power Kurtosis | Fourth statistical moment of the pulse rate power spectral density. | |
| Pulse Rate Power Peak | Spectral component of the pulse rate with the highest power. | |
| Pulse Rate DFA | Detrended fluctuation analysis of the pulse rate over 10 log-spaced windows, where the minimum window size is 10, and the maximum is 1000. A measurement of self-affinity. | |
| Pulse Rate AutoCorr 1 | First-order autocorrelation of the pulse rate. Measures the correlation of the pulse rate with a delayed copy of itself. | |
| Pulse Rate AutoCorr 2 | Second-order autocorrelation of the pulse rate. Measures the correlation of the pulse rate with a delayed copy of itself. | |
| Pulse Rate AutoCorr 3 | Third-order autocorrelation of the pulse rate. Measures the correlation of the pulse rate with a delayed copy of itself. | |
| Pulse Rate AutoCorr 4 | Fourth-order autocorrelation of the pulse rate. Measures the correlation of the pulse rate with a delayed copy of itself. |
Select a precomputed dataset to use for model training. Only datasets specifically configured for training are available. Cross-validation results are available during model testing.
Bayesian Optimisation
Training Progress
Bayesian Optimisation is used to search the hyperparameter ranges provided. This process attempts to balance exploration and exploitation to select the optimal values for the trained model. Gaussian Process Regression (GPR) and the Expected Improvement acquisition function are used.
For each set of hyperparameters, models are trained using the appropriate solver. K-fold cross-validation is used to produce a validation error which is input into the GPR that underpins the Bayesian Optimisation. The most optimal hyperparameters, as determined by the optimisation process, are then used to train the final model.
Models are calibrated using cross-validation data when training is complete so that predictions are accompanied by estimates of uncertainty. A trained Ridge Logistic Regression model is embedded within Gaussian Process and Support Vector Classifiers to produce a probability point-estimate during prediction, and a Laplace distribution is computed and embedded within Gradient Boosting, Linear, and Support Vector Regressors to produce confidence intervals that accompany the AHI point-estimate during prediction.
Trained models may be saved and are immediately accessible through the Analysis and Batch Analysis areas of the platform.
Select a precomputed dataset to use for out-of-sample Model Testing. Out-of-sample data are recommended.
Model Validation
Confusion Matrix
| Actual | ||||
|---|---|---|---|---|
| Positive | Negative | |||
Predicted |
Positive |
182 | 53 | PPV: 0.774 |
Negative |
114 | 433 | NPV: 0.792 | |
| Se: 0.615 | Sp: 0.891 |
Performance Metrics
Accuracy: 79%
Precision: 0.774
Recall: 0.615
F1: 0.685
AUC: 0.849
Positive = High Risk, Negative = Low Risk. PPV = Positive Predictive Value, NPV = Negative Predictive Value, Se = Sensitivity, Sp = Specificity, AUC = Area Under the Curve. For Gradient Boosting, Linear, and Support Vector Regressors, a predicted Apnoea-Hypopnea Index ≥5 is considered High Risk.
ROC Curve
Select a point to view the Threshold, False Positive Rate, and True Positive Rate.Summary of model performance generated by sweeping the classification threshold from high to low to produce a nonparametric Receiver Operating Characteristic (ROC) curve. Displayed in red is the current model operating point (threshold).
Model Validation data are collected from out-of-fold testing during cross-validation. These data are saved with the model and do not change.
Model Testing
Confusion Matrix
| Actual | ||||
|---|---|---|---|---|
| Positive | Negative | |||
Predicted |
Positive |
652 | 411 | PPV: 0.613 |
Negative |
189 | 2175 | NPV: 0.92 | |
| Se: 0.775 | Sp: 0.841 |
Performance Metrics
Accuracy: 82%
Precision: 0.613
Recall: 0.775
F1: 0.685
AUC: 0.891
Positive = High Risk, Negative = Low Risk. PPV = Positive Predictive Value, NPV = Negative Predictive Value, Se = Sensitivity, Sp = Specificity, AUC = Area Under the Curve. For Gradient Boosting, Linear, and Support Vector Regressors, a predicted Apnoea-Hypopnea Index ≥5 is considered High Risk.
ROC Curve
Select a point to view the Threshold, False Positive Rate, and True Positive Rate.Summary of model performance generated by sweeping the classification threshold from high to low to produce a nonparametric Receiver Operating Characteristic (ROC) curve. Displayed in red is the current model operating point (threshold).
Model Testing using the selected dataset. Data are out-of-sample and so testing results provide a reasonable indication of model generalisability.
Paired label and prediction data is available for download. Classification data are represented as Negative/Low Risk (-1) or Positive/High Risk (+1), while Regression data are available as an AHI.
Model Threshold
Set the model classification threshold. This is used for both classifiers and regressors to separate Wake or Sleep (Sleep Staging), or Low Risk and High Risk (Sleep Disordered Breathing) oximetry. Models must be saved before they can be updated.