Worldwide, Parkinson's disease (PD) is one of the most common neurodegenerative diseases afecting millions of people. Over the past 25 years, there has been a considerable evolution regarding Parkinson's diagnosis; initially, the only available tools were neurological examinations and autopsies. Subsequently, brain scintigraphy with DATSCAN was introduced and, in recent years, it has been added the magnetic resonance, an examination that doesn't use ionizing radiation, but only involves exposure to a magnetic field, capable of identifying neuromelanin's changes, visible as contrast reduction. In literature several studies have shown how impaired writing and vocal insuficiencies are important elements for the early detection of disease. In 2021, two studies by our research group on Parkinson's classification through telemedicine tools were carried out, one on graph signal (77.5The objective of our preliminary study is the development of a telemonitoring system of the voice and graph signal for neurological diseases' screening such as Parkinson's, thanks to artificial intelligence techniques (machine learning). The system is based on a simple Android/iOS application (on a smartphone or tablet), in which two tasks are required. The first one is vocal, and it consists of sounds recording, while the second one is related of writing letters and drawing geometric shapes. We enrolled 2 subjects, one healthy and the other one with advanced PD. Data are acquired via bluetooth wireless communication, stored locally, and sent to a web platform. Those data can be used to be processed both with standard analyses and through machine learning models (Artificial Intelligence) to support the specialist decision. Immediately afterwards, the two subjects underwent an MRI of the brain. The images are sent to the web platform and saved together with the data acquired through the application. Compared to previous research that used a professional tablet, a digital pen, and a microphone, our study introduced two novelties: the use of smartphones and tablets, and the archiving of radiological images on a platform. This system is a disease development control/monitoring tool that could save patients' time, allowing specialists to have all the patient's clinical and instrumental data on the platform, and having a positive impact on hospitals' resources. nd instrumental data on the platform, and having a positive impact on hospitals' resources.
It was observed the changing healthcare needs of individuals due to the progressive increase in the average age of the population, with a growing number of elderly people and chronic diseases. This requires a reorganization and structural redesign of the healthcare service network, with a focus on strengthening the territorial area of assistance. The increasing digitization we have seen in recent years is also making it possible to spread a range of new technologies that respond to the changing needs of healthcare professionals and patients. The efectiveness, eficiency, and appropriateness. At the European level, telemedicine is widespread in many countries (such as Sweden, Norway, Spain, and Great Britain) documents, and state projects in most cases. In Italy, the teleradiology, a telemedicine service that enables the elecNational e-Care Observatory was established in 2007 to tronic transmission of radiographic images from one geobuild the map of e-care networks, promote the exchange graphic area to another for reporting or consultation, can of good practices and related technologies, with the aim provide advantages, such as improved access to medical of improving the accessibility and efectiveness of on- expertise and reduced time to diagnosis and treatment. line services provided to citizens. Given the presence of numerous telemedicine initiatives that too often are not completed and remain incomplete projects, prototypes, 2. Materials and Methods or experiments, the Higher Health Council (in 2014, the agreement was signed in the State-Regions Conference), We consider two previous studies by our research group following a telemedicine working group started in 2010, that focused on the diagnosis of Parkinson’s disease usapproved the national guidelines for telemedicine in 2012, ing machine learning techniques, specifically in relation which will be periodically updated. The last national doc- to handwriting and graph analysis. The first study, called ument concerning telemedicine relates to the National ”Classification-Based Screening of Parkinson’s Disease Guidelines for the provision of telemedicine services, Patients through Graph and Handwriting Signals,” [1] resulting from the Permanent Conference for relations conducted in 2021 by Fratello M. et. al., aimed to propose between the State, the Regions, and the autonomous a method for discriminating and classifying Parkinson’s provinces of Trento and Bolzano in 2020 and will be pe- disease using artificial intelligence techniques [ 2]. The riodically updated based on new technologies and needs. study developed a telemonitoring system based on handIn Lombardy in 2020, the telemedicine regulatory sce- writing, which collected data on the inclination, pressure, nario was innovated with the entry into force of the and position of a digital pen during the execution of varDGR n. XI-3528/2020, of 05/08/2020, containing ”Indica- ious tasks by subjects [3]. The study collected data from tions for the activation of remotely deliverable healthcare 22 healthy individuals and 9 patients with Parkinson’s disservices (televisit).” With this act, the Region aimed to ease(PD), all of whom were right-handed, except for one provide uniform discipline and some specific operational patient with PD, and between the ages of 60 ± 25 years. indications for televisit. The Directive specifies that tele- The Hoehn and Yahr scale, a clinical scale used to devisit must be primarily employed in continuity of care scribe the progressive motor impairment of Parkinson’s (follow-up) for those patients already known to the health patients, was used to indicate the level of each patient. It service and who do not require an objective examination. was used a commercial Wacom One drawing tablet with Telemedicine, which is the use of technology to provide a screen for the test to extract both ”online”(pressed) and remote medical services, aims to improve the quality of ”ofline”(unpressed) features. The application used the healthcare, ensure continuity of care, increase access to Unity development platform to collect information on healthcare, and provide cost savings. Telemedicine in- the position (x, y), pressure, and inclination of the pen cludes various services such as teleconsultations with at a frequency of 133 Hz, and at the same time provided specialists, televisits, telemonitoring, and teleassistance. visual feedback on the tablet screen to the subjects. The Telemedicine is applied in diferent sectors of medicine Wacom tablets are widely used in the analysis of movesuch as teleradiology, telecardiology, teledermatology, tel- ment and handwriting because they ofer high spatial erhabilitation, telepathology, teleneurophysiology, and and temporal resolution. The study protocol was divided telehomecare. Various actors are involved in the organi- into four parts: drawing an Archimedean spiral, writing zation of telemedicine services, including users who can the bigram ”le” six times, writing two Italian phrases, be patients, caregivers, or medical professionals, and the drawing ten concentric circles, and writing seven lines provider center, which can be a National Health Service of free text. For each section of the protocol, a diferent facility, a general practitioner, or a pediatrician. The ser- screen was shown to the subject. The application had an vice center manages the health information generated initial page where the participant could enter their ID by the user and transmits the outcomes of the service to and a menu from which they could choose which activthe user. Wearable devices are biosensors integrated into ity to perform. The data was saved locally in diferent clothing, shoes, and accessories that measure biological “.csv” files for each acquisition, and MATLAB software parameters such as heart rate, respiratory rate, blood was used for analysis. During the execution of the tasks, pressure, glucose, and brainwaves. Wearable devices the participants were given visual instructions on the can provide feedback to the wearer and store data in the tablet screen, and the application recorded the data on cloud for healthcare professionals to access. Continuous the pen’s movements. We analyzed the data collected monitoring of biometric parameters can educate patients to determine if it could be used to diagnose Parkinson’s about healthy behaviors and modify their lifestyles to disease. The study showed that machine learning techpromote health and support diagnostic activities, treat- niques could successfully discriminate between healthy ment management, and rehabilitation. In particular the individuals and those with Parkinson’s disease based on
We present the results of our two studies focused on the development of machine learning models for the early diagnosis of Parkinson’s disease (PD) as in figure 1 . The Figure 1: Web platform permits raw data acquisition, feature ifrst study by Fratello M. et al. constructed three modtime variation visualization and machine learnings algorithms els to discriminate between healthy control subjects and implementation PD patients using data from two tasks: a spiral drawing task and a bigram task involving the repeated typing of the letter ”le.” The models were constructed using linear support vector machines (SVM) and k-nearest neighbor their handwriting movements. Overall, the study sug- (KNN) algorithms. The results showed that combining gests that handwriting analysis could be a promising tool data from both tasks led to the highest accuracy (77.5 for Parkinson’s disease diagnosis, and that telemonitor- per cent) and sensitivity (77.8 per cent), while the spiral ing systems based on digital pens and machine learning task alone led to the highest specificity (79 per cent). The techniques may provide a cost-efective and non-invasive best performing model used the medium KNN algorithm method for monitoring and diagnosing the disease. Go- for the combined tasks. The second study by Cordella ing on the research work, the new study involved 52 et al. focused on developing a machine learning model participants, including 16 patients with Parkinson’s dis- to classify PD patients and healthy controls based on ease and 42 controls, and was conducted using the same speech recordings. The authors tested several classifiers, Wacom tablet and digital pen. The participants were including SVM, KNN, and multilayer perceptron, on a asked to perform three tasks: drawing an Archimedean dataset of 612 observations. The results showed that spiral, writing the bigram ”le” six times, and writing the KNN had the highest accuracy (97.3 per cent) and the sentences ”I fiori sono sul prato” and ”Nel cielo ci sono le lowest standard deviation (less than 1 per cent). SVM had stelle” once each. The data collected were analyzed using low performance, and optimized SVM models resulted MATLAB R2022 software. The data analysis involved in a higher accuracy of 91.2 per cent but with a risk of applying an bandpass filter to eliminate non-relevant overfitting. The authors concluded that KNN was the sounds such as coughs, hesitations, and other non-vocal most robust classifier for this type of data. Overall, both sounds. The bandpass filter was applied to focus atten- studies suggest that machine learning models can be eftion on the 50Hz to 750Hz frequency range, where most fective tools for the early diagnosis of Parkinson’s disease. of the vocal signal exhibited its power. Amplitude and The findings also highlight the importance of selecting temporal thresholds were also applied, and all waveforms appropriate algorithms and data features for each task were normalized before being passed to the feature ex- and dataset [8]. The last results explore the possibility of traction algorithm. The features were divided into three using handwriting analysis as a non-invasive method to types: standard measures, non-standard measures, and diagnose Parkinson’s disease (PD). The study included cepstral measures. Standard measures consisted of con- 50 participants, 25 with PD and 25 healthy individuals. ventional measures of speech analysis such as jitter and The participants were asked to write a spiral, a sentence shimmer, while non-standard measures were based on with the word ”sono” (Italian for ”are”), a sentence with recurrence, self-afinity, and pitch dispersion. Cepstral the phrase ”i fiori sono sul prato” (Italian for ”the flowers measures were based on estimating the cepstrum. Dif- are on the lawn”), and a sentence with the phrase ”nel ferent types of models were tested on the data using cielo ci sono le stelle” (Italian for ”in the sky there are MATLAB’s Classification Learner and Weka. SVM, ANN, the stars”). The study used various machine learning aland KNN models were considered due to the need for gorithms to analyze the handwriting samples, including future firmware implementation. The models’ accuracy Subspace KNN, Medium KNN, Subspace Discriminant, was optimized by testing diferent subsets of features to Bagged Trees, Linear SVM, Fine KNN, Weighted KNN, analyze their robustness and prediction capacity. The Cubic SVM, Cubic KMM, and Subspace discriminant. The best models were selected for the validation session, in accuracy, specificity, sensitivity, precision, and area unwhich the data were divided into 75In addition to the der the curve (AUC) were measured for each algorithm. study’s results, the authors proposed an innovative idea The results showed that the handwriting analysis using to improve telemonitoring activity by creating a ”Voice machine learning algorithms had a higher accuracy than Drawing App” for smartphones or tablets that allowed for the previous study on the same topic. The best performdirect collection of handwriting and voice data without ing algorithms were SVM, KNN, and Tree. Additionally, the need for additional equipment [4] [5] [6][7]. the last study found that the models created with only the sentence ”i fiori sono sul prato”, the sentence ”nel cielo ci sono le stelle”, and the word ”sono” had the highest accuracy, possibly because they are located in the panded from 22 control subjects and 9 with PD to include central part of the sentence where the patients struggle the new data. Overall, the studies show the potential of the most, providing more information for the recogni- technology in aiding PD diagnosis through changes in tion of Parkinson’s disease. The study also suggested handwriting and voice patterns. However, more research a possible future work combining handwriting analysis is needed [10] to address limitations and to develop more and electromyography of the forearm to understand the robust algorithms that can be implemented in telemonirelationship between handwriting fatigue and disease. toring systems for accurate and reliable diagnosis. Finally, the study included a confusion matrix showing the performance of the models on the diferent sentences and the word ”sono”. References [1] M. Fratello, F. Cordella, G. Albani, G. Veneziano, 4. Conclusions G. Marano, A. Pafi, A. Pallotti, Classification-based screening of parkinson’s disease patients through We had presented three diferent studies related to using graph and handwriting signals, Engineering Protechnology to help diagnose Parkinson’s disease (PD) ceedings 11 (2021) 49. based on changes in handwriting and voice patterns. The [2] C. R. Pereira, D. R. Pereira, S. A. Weber, C. Hook, ifrst study discussed is by Fratello M. et al. (2021) and V. H. C. De Albuquerque, J. P. Papa, A survey on involves an application that records data from tablets computer-assisted parkinson’s disease diagnosis, at a frequency of 133Hz to help with the recognition of Artificial intelligence in medicine 95 (2019) 48–63. PD through changes in handwriting. The tool is simple [3] M. Thomas, A. Lenka, P. Kumar Pal, Handwritand easy to use, allowing subjects to complete the test ing analysis in parkinson’s disease: current status in the comfort of their own homes. The study collected and future directions, Movement disorders clinical data from 22 healthy subjects and 9 patients with PD practice 4 (2017) 806–818. and added it to the PaHaW database, which includes [4] A. Pompili, A. Abad, P. Romano, I. P. Martins, R. Cardata from PD patients and healthy control subjects. Us- doso, H. Santos, J. Carvalho, I. Guimaraes, J. J. Fering only two of the eight tasks included in the PaHaW reira, Automatic detection of parkinson’s disease: database, the study achieved an accuracy of 77.5 per cent, an experimental analysis of common speech prowhich is close to the 85.61 per cent accuracy achieved by duction tasks used for diagnosis, in: Text, Speech, Drotàr et al [9]. considering all eight tasks together. How- and Dialogue: 20th International Conference, TSD ever, the study’s main limitations are related to the small 2017, Prague, Czech Republic, August 27-31, 2017, number of subjects involved and the fact that the data Proceedings 20, Springer, 2017, pp. 411–419. was collected from diferent databases and in diferent [5] J. C. Vásquez-Correa, J. Orozco-Arroyave, T. Bockexperimental conditions, which could lead to incorrect let, E. Nöth, Towards an automatic evaluation of classification. The second study discussed is by Cordella the dysarthria level of patients with parkinson’s F. et al. and focuses on voice patterns. The study achieved disease, Journal of communication disorders 76 satisfactory performance in terms of accuracy (up to 98.5 (2018) 21–36. per cent) and stability through diferent combinations [6] M. Little, P. McSharry, E. Hunter, J. Spielman, of training. The accuracy had significantly lower fluctu- L. Ramig, Suitability of dysphonia measurements ations (<1 per cent) than in other studies, suggesting a for telemonitoring of parkinson’s disease, Nature more robust algorithm. However, due to an imbalanced Precedings (2008) 1–1. dataset with a majority of subjects with PD, sensitivity [7] F. Cordella, A. Pafi, A. Pallotti, Classification-based tends to be higher than specificity. The study shows the screening of parkinson’s disease patients through importance of analyzing complex sounds such as /iamh/ voice signal, in: 2021 IEEE International Sympoand /iuh/ and the ability of specific signal analysis to sium on Medical Measurements and Applications support medical work with a fast and reliable process (MeMeA), IEEE, 2021, pp. 1–6. that could be implemented in telemonitoring systems. [8] M. A. Hall, Correlation-based feature selection for The third study discussed is by Mancini A. et al., which machine learning, Ph.D. thesis, The University of collected data from 52 subjects (42 control and 16 with Waikato, 1999.
PD) who were asked to draw and write entirely in Italian [9] P. Drotár, J. Mekyska, I. Rektorová, L. Masarová, on a commercial tablet connected to a PC. This allowed Z. Smékal, M. Faundez-Zanuy, Evaluation of handfor the collection and saving of data. Like the Fratello et writing kinematics and pressure for diferential dial. study, this test was simple, quick, and comfortable to agnosis of parkinson’s disease, Artificial intelliperform, with healthy subjects taking three minutes and gence in Medicine 67 (2016) 39–46. those with PD taking four. The Italian database was ex- [10] A. Pallotti, G. Orengo, G. Saggio, Measurements comparison of finger joint angles in hand postures between an semg armband and a sensory glove, Biocybernetics and Biomedical Engineering 41 (2021) 605–616.