Diagnosis Support by Machine Learning Using Posturography Data
- 1. Diagnosis Support by Machine Learning Using Posturography Data Teru Kamogashira Department of Otorhinolaryngology and Head & Neck Surgery The University of Tokyo
- 2. Machine Learning • A method of data analysis that automates analytical model building. A branch of artificial intelligence based on the idea that systems can learn from data, identify patterns and make decisions with minimal human intervention. (sas.com) • Examples of practical use • Machine translation • Microsoft Translator, DeepL, Google Translate • Voice recognition • Cortana, Siri • Interactive system • The Microsoft Cognitive Toolkit, AI chatbot • Handwriting recognition • Facial image recognition • Medical imaging diagnosis • Automatic driving • Tesla
- 3. Methods of Machine Learning • Supervised learning • Data with labeled samples. • The desired output is known. • Regression • Classification • Unsupervised learning • Data has no historical labels. • The system is not told the “right answer.” The goal is to find some structures within. • Anomaly detection • Clustering • Dimensionality Reduction • Reinforcement learning • Discovers better solutions through trial and error - which actions yield the greatest rewards. • Often used in Robotics, Gaming and Navigation.
- 4. Differences between Statistics and Machine Learning • Statistical analysis • Data sets are small. • The model is easy to understand. • Analyzed by focusing on specific subgroups. • Machine learning • Data sets are large. • The model is hard to understand. • Operated automatically and systematically. • Analyzed by applying all features without excluding subgroups.
- 5. Posturography • Objective assessment of body equilibrium function. • Quantify the effect of vision with eyes closed (Romberg rate). • Evaluation of changes over time. • Not effective for the diagnosis of lesions. Eyes Open/ Eyes Closed With Foam Classic Patterns of Posturography Centralized Type Left-Right Type Anterior- Posterior Type Irregular Type Measures the trajectory of the center of gravity by measuring the ground reaction force generated when the body sways on the platform.
- 6. Prediction of Vestibular Dysfunction • For a better estimation of vestibular dysfunction • Increase parameters by load testing, etc. • Improve the analysis processing algorithm. • Use raw data directly with machine learning. Trajectory raw data Posturography Parameters ROC curve Prediction of Vestibular dysfunction Statistical Analysis Machine Learning Sensitivity 1 – specificity Area Velocity Length Amplitude Romberg ratio Frequency ... Direct processing Machine Learning
- 7. Increasing parameters by using Foam Posturography • Foam stress testing provides a more accurate estimate of the presence of vestibular disorders. Fujimoto C et al. Clin Neurophysiol 2009
- 8. Increasing parameters by using new methods • Frequency analysis with the maximum entropy method is effective. The green line indicates the borderline separating the presence or absence of vestibular disorders from the parameters of the gravitational sway test. MF AUC: The maximum entropy method / 0.1-1 Hz (middle-frequency range) AP: anterior-posterior direction LR: left-right direction Left: Conventional parameters Right: Using the maximum entropy method Fujimoto C et al. Otol Neurotol. 2014
- 9. Algorithm Improvement • To make better use of the existing data besides exploring new parameters. • Change the classification algorithm. • Use deep learning to utilize data thoroughly. • Pay attention to the overfitting problem. (taustation.com) Complex algorithms allow for more accurate classification.
- 10. Applying Machine Learning Algorithms to Postural Instability • Kamogashira T et al. Front Neurol. 2020 • Evaluated various machine learning algorithms in predicting peripheral vestibular dysfunction using a posturography test dataset. • Subjects • 75 patients with canal paresis evaluated with Caloric testing • 163 healthy subjects • Tested algorithms (in scikit-learn Python package) • Generalized linear model • Ensemble method • Support Vector Machines • Decision Trees • Multilayer Perceptron Classifier (Deep Learning) • Measures • K-fold cross validation • Recall • Area under the curve (AUC) of the receiver operating characteristic curve (ROC)
- 11. K-fold cross validation • Cross-validation is a resampling procedure used to evaluate machine learning models on a limited data sample. • “K” refers to the number of groups that the data is to be split into. (naokiwifruit.com) Split data into K groups
- 12. AUC of each Algorithm for Predicting Vestibular Dysfunction Some algorithms can make better predictions than classical methods from the same data. Deep learning does not always give good results.
- 13. Accuracy changes with Algorithm parameters • The accuracy varies depending on the parameters, and complex algorithms with high number or depth of trees do not always produce good results.
- 14. Summary of Results • The best algorithm • Gradient boosting classifier • AUC（0.90±0.06） Recall（0.84±0.07） • Classical algorithm • Logistic regression • AUC（0.85±0.08) Recall(0.78±0.07） • Multiple algorithms need to be evaluated on different clinical data sets. • The parameters of each algorithm need to be optimized to get better or best accuracy. • Deep learning does not always give good results.
- 15. Advantages and Disadvantages of Machine Learning • Advantages • Can be applied to a variety of problems. • Capable of greater accuracy and speed than humans. • Capable of processing large amounts of data that cannot be handled by humans. • Disadvantages • Need good data. • Fast processing may be required for learning. • Results can be difficult to interpret. • Difficult to adjust parameters and select features.
- 16. Towards Commercialisation • Valley of Death • Barriers between basic research prototype and commercial product. • River of Evil • Not enough performance. • Darwinian Sea • The product is not effective.
- 17. Toward Practical Applications in the field of Vertigo • The clinical data needs to be clean, garbage needs to be removed. • Database entry error. • Testing problems. • Dizziness symptoms differ between the early and chronic phases of the disease, which should be considered in the evaluation. • The goal needs to be clear. • Finding the disease that's in the mix. • Making an accurate diagnosis. • The more data, the better. • Better results when research is conducted collectively across institutions.
- 18. Thank you.
- 19. List of Evaluated Algorithms • Generalized linear model • Logistic regression • Stochastic gradient descent classifier • Ensemble method • Adaptive boosting classifier • Bagging classifier • Extra trees classifier • Gradient boosting classifier • Random forest classifier • Support Vector Machines • C-support vector classification • Nu-support vector classification • Decision Trees • Decision tree classifier • Extra tree classifier • Multilayer Perceptron Classifier (Deep Learning)