Inertial-sensor based 3D kinematics in the differential diagnosis between Parkinson's disease and atypical parkinsonism

• Autor(a) principal: Ferreira, Pedro Correia
• Data de Publicação: 2019
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositórios Científicos de Acesso Aberto de Portugal (RCAAP)
• Texto Completo: http://hdl.handle.net/10400.21/11785
• topic: Parkinson’s disease; Doença de Parkinson; Atypical parkinsonism; Parkinsonismo atípico; Differential diagnosis; Diagnóstico diferencial; Sensor-based kinematic analysis; Sensor baseado em análise cinemática; Applied machine learning; Aprendizado de máquina aplicado

Background and Objectives: Parkinson’s disease (PD) is the most frequent disorder presenting with Parkinsonism. However, atypical parkinsonian disorders (such as Progressive Supranuclear Palsy (PSP) and Multiple System Atrophy (MSA)) share some clinical features of PD but have different prognosis and are therefore important to distinguish from PD. The development of new and cost-effective tools to help clinicians in the differential diagnosis is therefore paramount. In this work, we perform sensor based kinematic analysis to find significant differences between PD and atypical parkinsonism, while developing an experimental machine learning model with clinical applicability. Subjects and Methods: 32 subjects with PD (mean age 69.7 ± 11.3, 14 male 17 female), 11 with atypical parkinsonism (9 PSP, 1 MSA, 1 vascular parkinsonism, mean age 72.9 ± 6.1, 8 male 3 female) and 33 age-gender matched controls (mean age 68.0 ± 12.6, 14 male 19 female) were recruited from the outpatient clinic in routine appointments at Hospital Egas Moniz, Lisboa. Using a set of 7 inertial sensors, leveraged by biomechanical models, we collected data from gait and posture during a 3x20m walk and stance. Moreover, using one inertial sensor, we recorded finger tapping tests. We conducted the analysis in two different ways: controls vs parkinsonian group (ill – PD + atypical), and within the parkinsonian group i.e. PD vs atypical. Results: Compared to controls, parkinsonian subjects displayed lower cadence (controls 109.420 ± 12.519 steps/min, ill 100.566 ± 13.432 steps/min, p = 0.006), step length (controls 0.512 ± 0.086 m, ill 0.442 ± 0.100 m, p = 0.003) and speed (controls 0.996 ± 0.190 m/s, ill 0.773 ± 0.202 m/s, p < 0.001). Double support was increased in the parkinsonian cohort (controls 36.355 ± 3.253 %, ill 39.827 ± 5.685 %, p = 0.003), and angular variables were decreased for the most part (i.e. hip flexion mean velocity controls 67.900 ± 14.974 cm/s, ill 55.180 ± 12.814 cm/s, p<0.001). A 10-fold cross validation random forest model classified controls vs parkinsonian subjects with an accuracy of 82.9%. PD and atypical cohorts also differed significantly, with the latter displaying high asymmetry in many angular parameters: knee mean velocity asymmetry (PD 0.919 ± 0.710 cm/s, atypical 3.392 ± 3.836 cm/s, p = 0.001), hip adduction mean velocity asymmetry (PD 0.268 ± 0.177 cm/s, atypical 0.656 ± 0.563 cm/s, p = 0.002) and ankle mean velocity asymmetry (PD 0.636 ± 0.537 cm/s, atypical 1.586 ± 1.227 cm/s, p = 0.002). The same machine learning model classified the PD vs atypical cohorts with 76.3% accuracy. Ongoing work is being developed in the analysis of finger tapping and postural metrics that can improve the discrimination models. Conclusions: The main objectives of this work were achieved. We hypothesize that atypical parkinsonian subjects develop an instable gait where asymmetry is highly pronounced compared to PD. Furthermore, we believe inertial sensor technology supported by machine learning should become a regularly applied technique in the differential diagnosis of these syndromes.