Machine Learning for Neurodegenerative Disease Diagnosis and Monitoring


A machine learning classification model incorporating biomarkers for detecting Parkinson's disease.

Key Benefits

  • The image processing methods are simple to use, and the processing pipeline is fully automated.
  • The method is novel and customized to address the practical requirements of clinical and research imaging.

Market Summary

Parkinson’s disease is a neurological disorder that causes uncontrollable movements, including tremors. Symptoms are often gradual to begin but worsen as the disease progresses. Over 10 million people globally are living with Parkinson’s disease and incidence increases with age. Close to 100,000 people are diagnosed with the disease in the United States each year. Current methods for diagnosing Parkinson's disease (PD) are limited and can be inaccurate. PD selectively damages pigmented catecholamine neurons in the locus coeruleus (LC) and substantia nigra pars compacta (SNc), and these structures undergo profound neurodegeneration in PD. However, most attempts at developing biomarkers for PD have been univariable and do not capture the complex and distinct multi-system biologies of the disease. Additionally, biomarker studies can be expensive and time-consuming to conduct.

Technical Summary

Emory researchers have developed a novel method to use multimodal MRI with specialized pulse sequences sensitive to specific facets of neurodegeneration combined with a customized machine learning approach to develop robust and novel multivariate biomarker profiles. The method involves acquiring measures using MRI pulse sequences sensitive to key features of neurodegeneration, including neuromelanin loss, iron accumulation, axonal degeneration, and microstructural damage. The multiple MRI measures within multiple ROIs across disease-implicated systems will comprise a high-dimensionality dataset from which to develop multivariate markers. The researchers then use a customized dimension reduction approach to select candidate measures to include in machine learning classifier models. This approach reduces the number of features in the machine learning model while increasing the likelihood that the included features will contribute the most to disease classification. The researchers used a machine learning approach, applying a logistic regression model to develop a multivariate classifier to differentiate PD patients from controls. They included SNc volume, LC volume, SNc R2*, and demographic features as inputs in the model. They limited the number of features included to these five and used 5-fold cross-validation to prevent overfitting. When assessed with receiver operating characteristic (ROC) analysis, the multivariate classification model had an AUC of 0.8557. In a sensitivity analysis, ANOVA comparing the relative importance of the features in the model found a statistically significant difference in importance among these biomarkers: age, gender, and three MRI measures (p-value < 0.001). The post hoc comparison found that the MRI measures, SNc volume, SNc R2*, and LC volume, were significantly more important than the demographic biomarkers. Applying the same approach in a second model, six clinical features of parkinsonian motor and non-motor symptoms assessed by self-report questionnaire along with six MRI features were included to classify PD and controls, and the observed AUC was 0.9418 in the ROC analysis for this model.1 A sensitivity analysis was again performed and two MRI features (SNc volume and SNc R2*) and two clinical features (MDS-UPDRS Part 2 questionnaire, REM sleep behavior disorder symptom questionnaire) were identified as highly informative to this model. This model performance is sufficiently high to be useful in both clinical diagnostic applications and in biomarker directed clinical trial designs. The clinical data included in the model are easily obtained remotely and do not require a physician to obtain.


  1. Langley, J., He, N., Huddleston, D. E., Chen, S., Yan, F., Crosson, B., Factor, S., & Hu, X. (2018). Reproducible detection of nigral iron deposition in 2 Parkinson's disease cohorts. Movement Disorders, 34(3), 416–419.
  2. Huddleston D*, Mahmoudi B, Langley J, Factor S, Crosson B, Hu X. MRI Signatures of Neuromelanin and Iron Pathology in Parkinson’s Disease. American Academy of Neurology Annual Meeting 2019. Philadelphia, PA, 2019 (Oral Platform Presentation).

Patent Information

Tech ID: 17225
Published: 8/31/2023