The Parkinson's Disease is one of many neurodegenerative diseases which slowly degenerates the central nervous system. It results from lack of dopamine in brain cells, and most often manifests itself in motor complications. The causes of such disorders have not fully been recognized yet. Thus, medical treatment of PD patients is limited to reducing the disease symptoms. The progress of the disease is slow and may take several years. In its early stage, the illness is hard to recognize, however, when diagnosed it must be effectively treated to reduce its further development. Main PD symptoms are motor complications such as bradykinesia, muscle rigidity, freezing of gait, tremor [1
], difficulties in swallowing, slow down or lack of facial expression and animation, etc. Beside causing motor disorders, PD may also impair concentration and daily routines planning.
Detecting the changes in PD patients state within short time is very difficult, in particular, because fully objective tests for evaluating the disease advancement have not been devised so far. One of the widely used methods to assess PD patients is conducting a series of normalized clinic tests known as the UPDRS--Unified Parkinson's Disease Rating Scale) [2
], which is a list of 42 items with numbers assigned from a 0-4 range. The list is divided into 4 parts [2
]. However, a subjective character of these tests introduces some error to their results.
Taken every three month or half a year, periodical UPDRS tests allow for the evaluation of the advancement of the disease. Such examination requires regular visits to specialists, and unfortunately this is often impossible to fulfill for PD patients for organizational reasons, for no access to specialists or simply because of deteriorating motor abilities. Other obstacles are problems with specifying the precise time of medications intakes or side effects of pharmacological therapy which cause dyskinesias (involuntary movement disorder). Some of late PD symptoms are consequences of long-term treatment with levodopa or dopamine receptor agonists [1
]. At the beginning, levodopa considerably improves patients' condition and may help maintain such state for a number of years. With time passing, the effectiveness of levodopa diminishes and patients start experiencing "on" (normal) and "off" (with parkinsonian symptoms) states alternatively. The changes from "on" to "off" states are evidently related to medication intake schedule, and they are predictable for "on" states. Some patients may, however, experience abrupt changes to "off" states with no correlation to the time they took medicine, and additionally a so called on-off phenomena of rapid changes from "on" to "off" states may appear [1
]. So, irregular or rare visits to physicians do not provide adequate information either on the advancement of the disease or on the appropriateness of the prescribed medication and its dosage. Above all, it is not possible to determine whether the patient is in the "on" or "off" state [1
]. This excludes a full objective evaluation of a PD patient's state.
PD patients' ratings achieved through clinical trials often base on historical data from patients diaries. However, these data are in most cases subjective and not sufficiently precise. Thus, during recent years, an increased interest in patients' objective assessment has been observed, and several attempts to predict the state of PD patients have been made. In particular, many studies show that wearable sensors provide a way to record human activities continuously. Human activity recognition (HAR) plays an important role in health and elderly people care [4
]. Not only the amount of movement but also precise information on its type is crucial in healthcare applications. Recognizing particular body activities let us detect disease symptoms, and analyze a patient's state [1
]. Typically, 3-axis accelerometers are used in HAR to capture movement characteristics for different body positions [10
]. The collected data can then be processed by different classifiers to recognize various activities of the monitored subjects. However, before applying any of the classification methods, a careful selection of pre-processing techniques and feature extraction methods should be performed.
A very valuable study by Patel et al. thoroughly reviews the state-of-the-art in the area of PD, and discusses limitations in PD patients' monitoring. The authors point out the main cause of limitations which for them is lack of integration between wearable technologies and algorithms used to estimate the severity of PD symptoms and motor complications. The paper by Patel et al. shows results of a pilot study on the feasibility of using data from wearable sensors to assess the severity of PD symptoms [1
]. In this study, the authors used the data collected by a set of sensors in comparison with video recordings captured during the examination. The clinicians evaluated UPDRS scores based on video recordings, and compared them with estimates derived from the accelerometer data [1
]. The paper also provides an exhaustive description of pre-processing and accelerometer-based feature extraction. The researchers used SVM (Support Vector Machine) with different configurations of settings and kernels as the classification algorithm. The paper by Patel et al. is one of the most important studies demonstrating that a continuous monitoring of PD patients can solve key problems in the assessment of PD progression. It also enables to estimate tremor, bradykinesia, and dyskinesia severity level [1
Home monitoring of PD patients via wearable technologies and web-based applications is another study of remote objective long-term health monitoring [14
]. Recently, the implementation of an iPhone estimating PD tremor with a wireless accelerometer application was also presented [15
]. Initial testing and evaluation of this application successfully proves its capability to acquire tremor characteristics in autonomous environments [15
]. A broader scope of healthcare application has a telemedicine instrument employed for remote evaluation of tremor: design and initial applications in fatigue and patients with Parkinson's Disease [16
Speech degradation is one of the early symptoms of PD. Tsanas et al. [17
] investigated tele-monitoring of PD progression by non-invasive speech tests. A methodology presented by Tsanas et al. is an example of mapping speech signal processing outcomes (e.g. dysphonias
which are malfunctions in voice production) to predict clinical overview utilizing UPDRS metrics [18
]. Two classification algorithms were employed in these studies--
the Classification and Regression Trees
(CART) and Random Forests
(RF). Their studies proved that the classifiers can replicate the clinicians' UPDRS estimates with the accuracy considered sufficient for PD assessment. Moreover, they provide a statistical evidence that speech impairment and average overall PD symptom severity are inherently related. Thus, the approach based on speech processing and classification may be justified for UPDRS progression prediction [21
A very promising approach was proposed by the PERFORM (A soPhisticatEd multi-paRametric system FOR the continuous effective assessment and Monitoring of motor status in Parkinson's disease and other neurodegenerative diseases progression and optimizing patients' quality of life) system, a European 7th FP project, that addressed this problem by designing and implementing a "Remote" Personal Health System [5
]. The system objective was to continuously monitor patients in their homes by recording selected motor and non-motor parameters, and data from specific accelerometer sensors, and passing them to clinicians and specialists at Central Hospital Units. The methodology behind the system was to capture symptoms of PD and automatically assign UPDRS ratings [2
] to them. After processing the data, PERFORM was supposed to generate alerts in the cases of emergencies [5
This paper is a continuation of the previous work performed within the framework of the PERFORM project. It differs, however, in the approach to data acquisition, as no data from sensors are used (see Figure ). Data come from patients' diaries and were rated by clinicians in the UPDRS scale. For the purpose of this study a computer application was prepared in Borland c++ environment. The training of the decision/support system was based on questionnaires used for classifying the motor state of the examined subjects. Decision tables were created and on this basis, a set of rules was generated taking into account historical and current UPDRS data from 47 subjects. This issue is described in the next Section. At the testing stage a decision algorithm that is incorporated into the system automatically assesses the patient's state. However to test overall efficiency of the decision system several algorithms were employed, first. Their efficiency and appropriateness for assessing the overall state of a PD patient are discussed in Section: Methods. Among tested algorithms the rough set-based approach was identified as the most efficient, and is applied in the methodology presented.
General scheme of the automatic assessment of the PD disease deteriorating.
In order to obtain medical knowledge, historical UPDRS data of 47 patients from Saint Adalbert Hospital in Gdańsk, Poland were gathered. All trials and investigations have been approved by the Ethical Committee of the Medical University in Gdańsk, PL. Also consents in a written form from all patients in the UPDRS examinations were obtained.
Table includes information on the subjects' sex and age. The subjects' average illness duration time was 9 (SD ± 5) years. The time period of historical UPDRS examination as compared to current examination was 8 months in average with the variance value of 7 months. For the assessment of PD progression, the UPDRS parts III and IV related to motor performance were used. Both historical and current evaluations were performed by clinicians. In the presented approach, the following 13 UPDRS items were assessed: UPDRS 13, 14, 20, 21, 23, 24, 25, 26, 28, 29, 31, 32, 39. Since PD is an asymmetrical disease, most of these symptoms are assessed separately for both right and left sides, resulting in 21 items.
Average and variance statistics of the subject used in the experiment
Five experts (neurologists) participated in the creation of the decision tables, but only four of them evaluated the current patients' data. Their task was to assign a criterion--a decision attribute ("stable", "worsening", "alert") to every possible pair of the given UPDRS item. Changes from the higher to the lower score were not evaluated. For each record in the decision table a histogram of experts' decisions was created. Examples of such decision tables and a histogram are presented respectively in Figures and . In the histograms information about number of experts voting for each criterion is presented (see Figure ).
An example of a decision table filled in by the clinicians.
Histogram related to a given decision table.
Since for 27 patients, an additional historical examination was available, in total 74 pairs of UPDRS evaluations between 'current UPDRS' and 'historical UPDRS' were used. As mentioned before the UPDRS pairs were assessed by 4 experts, thus the training set should consist in 296 training objects. However, eliminating superfluous data (entries in decision tables repeated) resulted in a reduced set of 284 elements.