The treatment of people suffering from Parkinson’s disease can be significantly improved through the use of digital techniques. How exactly artificial intelligence can contribute to an individual therapy is currently being analyzed in many studies and investigations. The aim is to be able to reliably detect even the first signs in order to be able to intervene in the course of the disease as early as possible.
Parkinson’s disease (PD) is the most common degenerative cause of parkinsonism. Atypical parkinsonism includes progressive supranuclear palsy (PSP), multiple system atrophy (MSA), and corticobasal degeneration (CBD). Clinical differentiation of parkinsonian syndromes remains difficult. Therefore, we investigated whether MRI and machine learning techniques improve diagnostic accuracy in patients with early-stage PD compared with clinical criteria [1]. 118 patients with suspected atypical parkinsonism in whom follow-up was available and brain MRI was performed at baseline were included in the study. Diagnoses at baseline and after a follow-up of two years were made using published clinical criteria. MRI diagnosis was based on radiological analysis of MRI images. T1-weighted images were segmented using FreeSurfer, an automatic segmentation software, and volumes of regions of interest were extracted: Midbrain, Pons, Cerebellum, and Basal Ganglia. A supervised machine learning algorithm (logistic regression), previously developed and trained with volumes of the same regions, was tested on this population. Subsequently, the diagnostic accuracy of the clinical criteria at the first visit, the radiological MRI analysis, and the machine learning algorithm were compared, using the final diagnosis as a reference.
The clinical diagnostic criteria were found to have a diagnostic accuracy of 63.6% at baseline. Radiological analysis of MRI correctly classified 82% of patients who met criteria for a possible diagnosis and 75% of patients with an unclear diagnosis at baseline. The algorithm also confirmed the diagnosis of parkinsonism in 91% of patients and in 66% of patients with an undetermined diagnosis. The results highlight the limitations of clinical criteria and the contribution of MRI to the early differentiation of parkinsonism. Although the accuracy was lower than that of MRI, machine learning could be of help in centers that are not experts.
Looked deep into the eyes
Neurodegenerative diseases (NDD) are the leading and increasing cause of disability worldwide. The increase in Parkinson’s rates is particularly alarming. Therefore, there is an urgent need for sensitive and specific biomarkers that allow differential prediction – especially in the early and prodromal phases of the disease, so that a personalized medical approach to therapeutics can be planned and pursued. The development of objective measures for Parkinson’s disease (PD) is complicated by the cognitive and motor spectrum of the disease, as well as by the presence of atypical disorders such as progressive supranuclear palsy (PSP). Eye tracking has been proposed as a source of prospective biomarkers in PD. Recent work has demonstrated the use of machine learning to classify PD and its cognitive spectrum based on oculomotor features. Now, increased sensitivity in an unstructured free-viewing task has been demonstrated for both PD and PSP [2].
120 PD patients, 8 PSP participants, and 97 age-matched control participants without neurological dysfunction performed a naturalistic free-vision task while their eyes were tracked with high accuracy. Saccade, pupil, and blink masses were extracted from the 10-minute movies. These measurements were used to train a classifier using a support vector machine. The classifier was adjusted and performance was measured using the area under the Receiver Operating Characteristic Curves (ROC-AUC) using a test series and cross-validation. PD and PSP were found to be predicted with high sensitivity using a free-view observation paradigm. The ROC-AUC was not only comparable to the antisaccade task, but even better. Next, the performance of this classifier will be evaluated using a naive test set by an independent body.
Classification based on brain scans
Future neuroprotective treatments for PD highlight the need for early diagnostic testing. MRI is not currently considered a robust imaging PD test. But exploratory techniques suggest that specific experimental sequences may be able to detect early pathological changes in the brain. Therefore, it was investigated whether such changes can be detected in routine MRI scans by using deep learning (DL) methods [3]. This subset of machine learning has recently shown great promise for diagnostic medical imaging, as it has the potential to detect patterns invisible to the human eye. New explanatory methods make it possible to better interpret DL predictions.
194 scans were taken more than four years after diagnosis, 265 two to four years after diagnosis, 241 one to two years after diagnosis, and 282 less than one year after diagnosis. Each cohort was matched to controls based on age and sex. The longer the time since diagnosis, the better the diagnostic performance of the models. The models trained on the later cases of PD showed good diagnostic performance. Declining performance for earlier stages of PD suggests that progressive changes have been identified. The use of explainable AI has highlighted regions consistent with the known neuropathology of PD and provides a focus for future work.
Detect and classify tremor
The main symptoms of PD are bradykinesia, tremor and rigidity. PD diagnosis depends primarily on clinical assessment using the Movement Disorder Society-sponsored revision of the Unified Parkinson Disease Rating Scale (MDS-UPDRS). One study attempted to develop a machine learning program for tremor detection and classification in patients with Parkinson’s disease [4].
A triaxial mechanical accelerometer was developed to objectively detect tremor in PD and related movement disorders in addition to clinical assessment. For this purpose, a low-cost quantitative continuous measurement of movements in the extremities of people with PD – a modification of the MDS-UPDRS – was used. The protocol was performed by trained MDS-UPDRS certified assessors on 20 participants with PD and eight age- and sex-matched healthy controls. Ten-second segments of acceleration signals of repetitive tasks (finger tapping, hand movements, pronation-supination of hands, toe tapping, and leg mobility) and their Fast Fourier (FFT) and Continuous Wavelet Transforms (CWT) were classified into two classes: low (corresponding to ratings 0-1) and high (corresponding to ratings 3-4). An equal number of images were randomly selected from each class and used for classification. The ability of the network to correctly classify the validation images determined its hit rate.
The network had 92% accuracy in predicting new CWT images and 97% accuracy in predicting new STFT images in low (0-1) and high (3-4) classes. Using machine learning to categorize the output of movement instruments is a viable technique for classifying repetitive movements of people with PD and healthy controls. The experiment will be repeated with 100 patients and 100 healthy controls.
Targeting tics
Distinguishing between tics in people with tic disorders and additional movements in healthy controls can be difficult. In addition, evaluating tics from video recordings is time-consuming and tedious. Machine learning has the potential to help with these challenges by distinguishing between tics and other additional movements and supporting clinical assessments.
In one study, a dataset of 63 videos of people with tic disorders was used to train a random forest classifier for second-by-second tic detection [5]. The classifier used facial features as input and defined tic seconds as those with tics equal to or greater than a predefined threshold. The trained classifier was then used to predict the presence of tics in patients and additional movements in healthy controls. These predictions were used to calculate various features, such as the number of tics per minute, the maximum duration of a continuous non-tic segment, the maximum duration of a continuous tic, the average duration of tic-free segments, the number of changes from tic to non-tic segments and vice versa per minute, the average size of a tic cluster, and the number of clusters per minute. These characteristics were combined into a single tic recognition score using logistic regression. Model parameters were obtained by training with a dataset of 124 videos of individuals with tic disorders and 162 videos of healthy controls. To evaluate the accuracy of this score in classifying patients and healthy controls, a test data set of 50 videos from patients and 50 videos from healthy controls was used. The test set achieved a classification accuracy of 83%.
The machine learning algorithm is useful for detecting tics and distinguishing between tics and other additional movements. It could be developed into a clinically applicable tool. To improve classification accuracy, the next step is to fine-tune the score for tic detection. Furthermore, the importance of each characteristic will be analyzed to determine which characteristics are most helpful in distinguishing between the two groups. The algorithm could then also be helpful in distinguishing between tics and functional movements.
Congress: International Congress of Parkinson’s Disease and Movement Disorders® 2023
Literature:
- Chougar L, Faucher A, Faouzi J, et al.: Contribution of MRI and machine learning approaches to the diagnosis of patients with early-stage parkinsonism in a situation of clinical uncertainty [abstract 161]. Mov Disord 2023; 38 (suppl 1).
- Brien D, Riek H, Ye R, et al.: Machine learning classifies Parkinson’s Disease and Progressive Supranuclear Palsy on saccade, pupil, and blink measures during a naturalistic free-viewing task [abstract 276]. Mov Disord 2023; 38 (suppl 1).
- Courtman M, Thurston M, Mcgavin L, et al.: Explainable deep learning based detection of Parkinson’s changes in MRI brain scans [abstract 1552]. Mov Disord 2023; 38 (suppl 1).
- Elshourbagy T, Hernandez M, Mckay G, Brasic J: Artificial Intelligence to detect and classify tremors in patients with Parkinson’s disease and related conditions [abstract 1211]. Mov Disord 2023; 38 (suppl 1).
- Becker L, Schappert R, Brügge N, et al.: New machine learning approaches in tic detection: Seeking to learn more about the characteristic of tics [abstract 951]. Mov Disord 2023; 38 (suppl 1).
InFo NEUROLOGIE & PSYCHIATRIE 2023; 21(5): 24–25