Ifmbe proceedings 29 - an automated method for levodopa-induced dyskinesia detection and severity classification

An Automated Method for Levodopa-Induced Dyskinesia Detection and
Severity Classification
M.G. Tsipouras1, A.T. Tzallas1, G. Rigas1, P. Bougia1, D.I. Fotiadis1 and S. Konitsiotis2 1Unit of Medical Technology and Intelligent Information Systems, Department of Material Science and Engineering, University of Ioannina, 45110 Ioannina, Greece 2 Department of Neurology, Medical School, University of Ioannina, 45110 Ioannina, Greece Abstract— In this paper we propose an automated method
ology several computer-based methods are developed using for Levodopa-induced dyskinesia (LID) detection and classifi-
quantitative instrumental techniques such as: movement cation of its severity. The method is based on the analysis of
sensors (accelerometers and gyroscopes) [2-6], electromy- the signals recorded from accelerometers which are placed on
ography (surface) [2,7], force gauges (which are instru- certain positions on the patient’s body. The signals are ana-
ments used to measure the force during a push or pull tests) lyzed using a moving window and several features are ex-
tracted. Based on these features a decision tree is used to detect
[2,8], position transducers (force transducer that measured if LID symptoms occur and classify them related to their se-
arm movements) [2,8] and Doppler ultrasound systems verity. The method has been evaluated using a group of pa-
[2,8,9]. Methods which are based on accelerometer signal tients and the obtained results indicate high classification
analysis, greatly differ in the body segments, from which ability (95% classification accuracy). Furthermore, extensive
movements are measured, and the number of accelerometers evaluation has been done in order to determine the optimal
per segment. A major challenge for a method that automati- positioning of the sensors and the selection of the classification
cally assesses LID is to be able to distinguish dyskinesias algorithm.
from voluntary movements. Most of the studies tried todetect LID focused on the frequency domain of the signals Keywords—
Levodopa-induced dyskinesia
detection,
Levodopa-induced dyskinesia severity classification, auto-
from the movement sensors, while time-domain features mated diagnosis.
have also been used. Severity of LID has been determinedusing linear discriminate analysis and artificial neural net-works (ANNs). An important drawback of the aforemen- tioned studies is the small number and the short-time oftasks involved as well the fact that they have been per- Levodopa-induced dyskinesia (LID) is a disabling and formed in laboratory settings. Keijsers et al. [6] monitored distressing complication of chronic levodopa therapy in patients while performing a large variety of daily life activi- patients who suffered from Parkinson's disease [1]. LID is ties in a natural environment for a long period of time. Hoff more commonly known as a jerky, dance-like movement of et al. [10] use a continuous ambulatory multi-channel accel- the arms and/or head. These movements (called as choreic erometry (CAMCA) to identify accelerometer characteris- or dystonic) range from 1-5 Hz [1,2]. LID symptoms can be rated in various ways by their topography (affected body In this study we propose a method for LID detection and regions), by their duration or consistency of effect, by the classification of its severity. Six accelerometers are placed disability they impart, by the extent of enhanced severity on the patient’s body and the recorded signals are analyzed from activation due to volitional movement and by the se- in order to extract several features. The analysis is per- verity [2]. The presence and severity of LID can change formed using moving windows. All features extracted from during the day and as such, detection, assessment and fol- a specific window of the signal for all signals (from differ- lowing the changes of these signs during daily activities are ent sensors) form a feature vector that is used to detect if of great interest. In addition, the effective characterization LID symptoms which are present on this window and de- and quantification of LID not only improves our under- termine their severity. The classification technique that is standing of its pathophysiological mechanisms, but also employed is a decision tree. Several experimental settings helps diagnosis and the evaluation of treatment.
related to the number of sensors used have been evaluated Current assessment of LID mainly relies on clinical me- and the results are presented. In addition, based on the best thods [2,3]. Unfortunately, clinical methods lack objectivity experimental settings determined from the above analysis, and they are not feasible for long-term assessment by the other classifications techniques are also tested and the ob- experts [2,4]. To overcome the limitations of subjective assessment of LID and to gain insight into their pathophysi- P.D. Bamidis and N. Pallikarakis (Eds.): MEDICON 2010, IFMBE Proceedings 29, pp. 590. www.springerlink.com An Automated Method for Levodopa-Induced Dyskinesia Detection and Severity Classification three signals, one for each axis (x,y and z axis). The abovesensor’s placement on the patient’s body is illustrated inFig. 1. All sensors transmit data using Bluetooth to a porta- ble PC equipped with data acquisition hardware and soft- In this study three patients, two males (aged 65 and 75 ware to collect and store the signals. The sensors’ size is no years) and one female (aged 60 years) were enrolled. They bigger than a matchbox. Sensors on the arms and legs are suffered from LID and showed a severity of LID varying attached on specially designed elastic bands which allow between no dyskinesia to moderate (rating between 0 and 3 fixation to any wrist or ankle size. Sampling rate is set to on the Unified Parkinson’s Disease Rating Scales (UPDRS) [11]. The experiments were approved by the Medical Ethi-cal Committee of the Hospital of the University of Ioanninain Greece.
Following a standard procedure, used also in clinical tri- als with medication which accepted in the literature, pa-tients should have received the last dose of their medica-tions 12 hrs before testing time, which is usually around 8pm of the night before. Twelve hours after the last dose ofmedication the patient is expected to be in the “off” state.
Recording started with the patient always being in the “off”state and lying on his bed. The protocol consists of threemajor tasks: Fig. 1 Schematic overview of the position of accelerometers on the patient’s body.
x rising from the bed and sitting on a chair located just x standing up from the chair and performing a series of activities (for totally of approximately 8 min): walking The recorded signals are used for feature extraction. A for a distance of 5 m, opening a door, closing the door, moving window with 2 seconds duration and 1.75 seconds opening the door step out of the room, walking in the antepossition is used over a single lead and, for each win- corridor for a straight distance of 10 m, returning in dow the mean value of the signal is calculated: the room, making a stop, drinking a few sips from a glass of water, returning to the chair.
Then, the patient takes his first dose of medication for that ݅ ݆) is the mean value of the ݆ݐ݄ window of the ݅ݐ݄ day and when he turns “on” (verified on site by an expert signal, ݂ݏ is the sampling frequency, ݓ(݆) is the time posi- neurologist), another cycle of recording with the above prespecified tasks follows. If the patient had LID (while in ݅ݐ݄ signal. The above procedure is applied to each recorded the “on” state) then the recording is selected for this study.
signal. Then, very slow movements are modeled: UPDRS obtained immediately before the patient started performing the predefined tasks. The final annotation re- lated to the LID severity based on UPDRS is made based onvideo recordings obtained during the protocol procedure and subsequently subtracted from the recorded signals: A moving window with 2 seconds duration and 1.75 seconds anteposition is used over a single lead and for each The movements and postures are measured using accele- window the standard deviation of the signal is calculated: rometers and a portable data recorder. Six sets of three or- thogonal accelerometers (ANCO Devices [12]) are used.
These are placed at six different positions of the body: rightand left wrist (LW and RW), right and left leg (LL and RL), ݅ ݆) is the standard deviation of the ݆ݐ݄ window of chest (CH) and waist (WS). Each accelerometer records ഥ݅ is the corresponding mean value.
Table 2 Classification Accuracy (%) of the Experimental Settings For each window a feature vector is created. This vector the dimension of the feature vector is 2 כ 3ܰ, where ܰ is the number positions on the patient’s body (for each posi- tion three signals are recorded). This feature vector is used for LID assessment using a decision tree. Decision trees are a widely used classification technique. They represent theacquired knowledge in the form of a tree. The tree can be easily transformed to a set of rules with mutually exclusive The construction of the decision tree is implemented us- ing the C4.5 inductive algorithm [13]. This algorithm con- structs a decision tree from the training data. Each internal node of the tree corresponds to a principal component, while each outgoing branch corresponds to a possible range of that component. The leaf nodes represent the class to beassigned to a sample. The C4.5 algorithm applies to a set of data and generates a decision-tree, which minimizes the expected value of the number of tests for the classification Based on the signal analysis described above a classifica- tion dataset was formed. The number of instances related to the patients and the LID severity are shown in Table 1.
Table 1 The Dataset used in this Study.
other classifiers are evaluated. This include Naive BayesClassifier (NBC), k- Nearest Neighbour (k-NN), Fuzzy Lattice Reasoning (FLR [14]), Decision Trees (C4.5) and Random Forests (RF [15]). The results in terms of classifi- cation accuracy (%) are presented in Table 3.
Table 3 Classification Accuracy (%) for the 21 and 22 Several different experimental settings have been used, related to the combination of signals which are used for LID assessment. For each one of them, results are obtained in terms of sensitivity, specificity and classification accuracy.
The 10-fold stratified cross validation is used in all cases.
The various combinations of signals used in each experi- mental setting and the obtained classification accuracy are Additionally, for the final two experimental settings i.e.
the hands, legs and chest sensors (setting 22 in Table 2) and hands, legs and waist sensors (setting 23 in Table 2), An Automated Method for Levodopa-Induced Dyskinesia Detection and Severity Classification A method for the automated LID detection and classifica- Keijsers NL, Horstin MW and Gielen SC et al (2003) Online Moni-toring of Dyskinesia in Patients with Parkinson’s disease. IEEE Eng tion of its severity based on the analysis of signals obtained by accelerometers placed on the patient’s body is presented.
Hoff JI, van Hilten BJ and Roos RA (2001) A review of the assess- The method has been evaluated using recordings from three ment of dyskinesias. Mov Disord 14(5): 737 – 743 patients that presented LID severities 0 to 3 at the UPDRS.
Keijsers NL, Horstin MW and Gielen SC (2003) Movement parame-ters that distinguish between voluntary movements and levodopa- The features extracted from the signals carry sufficient induced dyskinesia in Parkinson's disease. Hum Mov Sci 22(1): 67-89 information for the LID severity detection and classification Burkhard PR, Shale H, Langston JW and Tetrud JW (1999). Quantifi- since they are the local mean value, which is related to very cation of dyskinesia in Parkinson's disease: validation of a novel in- slow dystonic movements, and local standard deviation, strumental method. Mov Disord 14(5):754-63 Keijsers NL, Horstink MW, van Hilten JJ, Hoff JI and Gielen SC which depicts the faster jerky, dance-like movement of the (2000) Detection and assessment of the severity of Levodopa-induced limps and/or head. LID effects may be present in a single dyskinesia in patients with Parkinson’s Disease by neural networks.
part of the patient’s body (i.e. only one hand) or to several (i.e. both hands and head). Also, the effects may be present Keijsers NL, Horstink MW and Gielen SC (2003) Automatic assess-ment of Levodopa-induced dyskinesias in daily life by neural net- in the limps (hands/legs) or affect the whole body (waist/chest). Thus, the feature vector included accelerome- Yanagisawa N (1984) EMG characteristics of involuntary move- ter signals from several positions of the patient’s body. The ments. In: Dyskinesias. Bruyn GW, Ed. Sandoz BV, Uden, 142-159 obtained results indicate that the proposed method is highly Xuguang Liu, Carroll CB, Wang SY, Zajicek J and Bain PG (2005)Quantifying drug-induced dyskinesias in the arms using digitized spi- efficient for automated LID severity detection and classifi- ral-drawing tasks, J Neurosc Meth 144(1): 47-52 Haines J and Sainsbury P (1972) Ultrasound system for measuring The results presented in Table 2 indicate that experimen- patients' activity and disorders of movement. Lancet 14:2(7781):802- tal settings that include signals from almost all positions of 10. Hoff JI, van den Plas AA, Wagemans EA, and van Hilten JJ (2001) the patient’s body present the best results. However, this Accelerometric assessment of levodopa-induced dyskinesias in Park- conclusion was anticipated since (as mentioned earlier) LID effects may be present to a single or to several parts of the 11. Fahn S, Elton R (1987) UPDRS Development Committee. Unified patient’s body. This also confirms that in our dataset the Parkinson’s Disease Rating Scale. In: Fahn S, Marsden C, Calne D,eds. Recent Developments in Parkinson’s Disease. Florham Park, NJ: presence of LID effects to the patient’s body is time varying Macmillan Health Care Information, 153–164 i.e. the same patient may presents LID effects in different parts of his body for different time intervals. Thus, a method 13. Quinlan JR (1993) C4.5: Morgan Kauffman California that is based on a selection that includes signals from sev- 14. Hoff JI, van den Plas AA, Wagemans EA, and van Hilten JJ (2001) Accelerometric assessment of levodopa-induced dyskinesias in Park- eral position of the patient’s body (such as experimental settings 22 and 23) is expected to present the best results 15. Fahn S, Elton R (1987) UPDRS Development Committee. Unified (compared to selections that include a limited number of Parkinson’s Disease Rating Scale. In: Fahn S, Marsden C, Calne D, eds. Recent Developments in Parkinson’s Disease. Florham Park, NJ:Macmillan Health Care Information, 153–164 In this study, the classification technique that is selected is a decision tree based on the C4.5 algorithm. The results presented in Table 3 indicate that a selection of a more Institute: Department of Material Science and Engineering, Universi- advanced technique, such as random forests, does not im- prove significantly the obtained results.
This work is part funded by the ICT program of the Eu- ropean Commission (PERFORM Project: FP7-ICT-2007-1-215952).

Source: http://www.qbase.gr/sites/default/files/27%20-%202010%20MEDICON(12).pdf