A holistic approach to assess the feasability of telehealth for Parkinson's disease management

This work of thesis has been entitled “A holistic approach to assess the feasibility of telehealth for Parkinson’s disease management” since it is aimed at exploring the feasibility of telehealth, understanding it as the use of technology and related services to deliver a disease management at distance, to provide a tool for the management of Parkinson’s disease (PD). The term holistic has been incorporated in the title since it tried to cover not only the feasibility of developing the appropriate technology for the continuous monitoring and assessment of the disease remotely and in unsupervised environment, i.e. at patients’ homes, but also because it covers the feasibility study from a wider perspective by incorporating the analysis of human factors and Health Technology Assessment (HTA). These two topics are aimed at incorporating the perspective and views of the patients and health experts and provide data for policy-makers, health care managers but also to the developers and researchers about the potential uptake of telehealth for Parkinson’s disease in the current clinical practice. This work was framed within the European project PERFORM, on this research project a Body Area Network (BAN) of wearable sensors was designed and developed for the automatic detection and quantification of the motor symptoms of PD which was then transmitted to a hospital platform also developed within the project. There, the motor assessment was analysed together with other sources of information also collected with the PERFORM system at patients’ homes such as food and drug intakes. Finally, all this information provided clinicians with suggestions about treatment modification when needed. On this context, the research work presented on this thesis is divided into three independent parts. The first part covers the use of the raw data collected with the BAN for the automatic assessment of the bradykinesia and the gait disturbances. On this regards, two different methodologies that have been previously explored on supervised environments were explored to be translated in the context of a home telehealth platform. In the gait case, the inverted pendulum model was used to estimate different gait features such as velocity, step frequency and step length. Other metrics like arm swing and entropy were also extracted. The feasibility study tried to explore whether the proposed BAN was feasible for the implementation of such algorithms. For the bradykinesia case, a different non-linear approach was followed, specifically machine learning has been used, exploring diverse configurations in order to find about the best setup for bradykinesia assessment, then a meta-analysis methodology was proposed to improve the outcomes of the machine learning algorithms. The results presented for the bradykinesia assessment showed the best outcomes taking the signals coming from wrist sensor as input while the subject was performing some hand movement, on this case the accuracy of the classifier achieved up to 86.48% of correct classified instances. The accuracy in the bradykinesia severity detection while the subjects were walking was slightly lower (ranging from 70.83% to 75%). In all the cases, the combination of the rms+range features showed the better performance. Using the proposed meta-analysis method an average of 74.4% of agreement between the clinical examination and the system outcomes was achieved, ranging from 91.9% to a 48.8%. To study the gait performance, SampEn was evaluated as a metric to characterize the degradation of the gait performance confronting PD patients and healthy subjects, the values of SampEn were significantly different and provide useful information especially in the waist sensor (p<0.01) and the left arm (p<0.01), right arm (p=0.05). Regarding the estimation of the gait features for PD, the mean error of Step frequency, Stride Length and Speed, calculated on 30 trials and 10 subjects are respectively: 3.36%±2.57, 7.31%±4.69, and 7.77%±5.34. The average data loss satisfied that in all cases more than 95% of the data is successfully transmitted. Nevertheless, analysing the average burst length and the standard deviation of these values we found burst of up to 143.04±240.10ms. A burst of these characteristics in the middle of the transmission would compromise seriously the performance and accuracy of the system. Nevertheless, adding the most frequent burst length as an extra measure and observing the raw data we can easily identify that 64ms is the most common burst in all the cases, moreover the longest bursts are always at the beginning of the session i.e. during the start-up process. That means that waiting for all the sensors to be properly connected before start computing the gait features will remove the longest burst errors. The second part of this thesis addresses the human factors related to the use of a telehealth platform for PD. Particularly, the work encompassed on this thesis describes a methodological approach to extract and analyse the PD patients feedback regarding the wearability of the PERFORM system. A combination of several methodologies already validated have been emplaced, specifically the REBA, Borg and CRS scales have been used in combination with a body map when needed. A total of 32 PD patients have been involved in this study showing an excellent acceptance of the PERFORM system in wearability terms. The overall result of the evaluation can be broken down in the subsequent marks: average perceived exertion score (0.71 on the Borg CR-10 scale), average posture score (0 based on the REBA scale), average localized pain/discomfort score (0.11 on the Borg CR-10 scale), average emotion score (1.4 on the CRS scale), average attachment score (2.5 on the CRS scale), average harm score (0.15 on the CRS scale), average perceived change score (0.67 on the CRS scale), average movement score (0.59 on the CRS scale) and average anxiety score (1.63 on the CRS scale). Each individual component of the wearability assessment can be labelled as low level of effect according to the guidelines provided by each of the methodologies. Finally, the last part of the thesis is aimed at identifying tools available and viable to carry out an early HTA on telehealth systems for Parkinson’s disease (PD) management and apply such methodologies to assess the PERFORM system as a case of study. The whole chapter is divided in two parts addressing two complementary works which try to provide two complementary points of view. The first part is aimed at identifying the elements that a telehealth system for PD management should satisfy to be effective and aligned with the users’ needs as well as to understand if clinicians and technicians share a common opinion when evaluating these systems. Instead, the second part provides a methodology to estimate the potential economic impact, in terms of cost-effectiveness, of a real implementation of telehealth systems for PD management. After the exploration of the available tools for early HTA two methodologies were selected. The first one, called Analytic Hierarchy Process (AHP), allows the elicitation of the user needs and the definition of a framework for the evaluation of telehealth solutions for Parkinson’s disease patients. 16 experts in the domain both with clinical and technical background were involved in the study. The second part of the chapter explores how the use of Markov models with a minimal number of states were used to perform “what-if” and finally, to simulate and estimate potential scenarios for the economic impact, in terms of cost-effectiveness, of a telehealth platform. Then, this approach was used to estimate the potential cost-effectiveness of the PERFORM project. Finally, a sensitivity analysis will be carried out. According to the results of the responders, the three most important user needs are Increase the wearability acceptance (0.102), Increase the self-management support (0.093) and On/Off fluctuations detection (0.085), based on the values from the Global Weights (GW). In a second level of importance were found Increase patient & carers knowledge (0.080), Assist care givers (0.070), User-friendly interfaces (0.068), Motor symptoms assessment (0.064) and Increase patient Quality of Life (0.063). The agreements between the responders of the technical and clinical groups showed a significant correlation between both, ρ = 0.590 (p=0.012). Regarding the use of Markov modelling, six general “what-if” scenarios were simulated providing a general overview of how inputs variations could impact the ICER values. These scenarios provide a quantitative reference framework for any health technology related to Parkinson’s disease and could be useful to understand and to orientate the research, particularly in early stages of development, towards an efficient cost-effectiveness path. Then, the PERFORM system was used as case of study with the Markov analysis. It was confronted with the base case and the cumulative costs and QALYs were calculated for each one and the ICER was calculated at year 5. The ICER values estimated by the model are 6397.36 for the societal perspective and 8977.20 for the third party perspective. These values suggest that PERFORM will not be a cost-saving technology (positive ICER) but these values are below most of the ICER thresholds commonly accepted worldwide (10,000€/QALY). Finally, by placing the PERFORM innovation from the societal and third party perspective in the ΔQALY-ΔC plane it was possible to see the margin in incremental cost to the different ICER thresholds. Taking the 30,000€/QALY as reference, the PERFORM system has a margin of 20855.28€ incremental cost in the societal perspective and 18575.73€ incremental cost in the third party perspective. A sensitivity analysis was also performed to explore the robustness of the results and the limitations of this approach were properly identified. Resumen Este trabajo de tesis ha sido titulado “A holistic approach to assess the feasibility of telehealth for Parkinson’s disease management” ya que pretende abordar la viabilidad de utilizar una plataforma de telesalud, definiendo esta como el uso de la tecnología y los servicios asociados a ella para la gestión de una enfermedad de forma remota, con el fin de proveer una herramienta que permita la gestión de la Enfermedad de Parkinson (EP). El término holístico ha sido incorporado en el título ya que este trabajo pretende cubrir no solo la viabilidad del desarrollo de la tecnología apropiada para la continua monitorización y evaluación de la enfermedad de forma remota y en un entorno no supervisado, esto es, en la propia casa de los pacientes, sino que también pretende estudiar la viabilidad desde una perspectiva más amplia incorporando un análisis de factores humanos y Health Technology Assessment (HTA). Estos dos temas pretenden incorporar el punto de vista de los pacientes y los expertos y proporcionar la información necesaria a los responsables de la elaboración de políticas y regulaciones, así como a los gestores, desarrolladores e investigadores sobre el potencial de dicha tecnología para ser realmente incorporada a la práctica clínica. Este trabajo está enmarcado dentro del proyecto europeo PERFORM. En dicho proyecto, una Body Area Network (BAN) de sensores vestibles fue diseñada y desarrollada para la detección automática y cuantificación de los síntomas motores de la EP, que después eran enviados a un módulo instalado en el hospital y también desarrollado dentro de dicho proyecto. Allí, la evaluación motora era combinada con otras fuentes de información también recogidas con el sistema PERFORM en la casa de los pacientes como, por ejemplo, información sobre las comidas y los medicamentos. Finalmente, la plataforma para los neurólogos proporcionaba sugerencias acerca de posibles cambios en el tratamiento utilizando un motor de reglas y en función de la información recogida. Basado en esto, el trabajo contenido en esta tesis se puede dividir en tres partes. La primera parte cubre el uso de las señales recogidas con la BAN para la evaluación automática de la bradiquinesia y alteraciones en la marcha. Dos metodologías, previamente validadas en entornos supervisados, fueron empleadas y la viabilidad para trasladarlas a entornos no supervisados fue estudiada. En el caso del estudio de la marcha, se utilizó el modelo del péndulo invertido para estimar diferentes características de la marcha como, por ejemplo, la velocidad, frecuencia de paso y la longitud de paso. Otras métricas como la frecuencia de braceo y la entropía de la señal también fueron calculadas. El estudio de viabilidad se centró en analizar si la BAN propuesta podía calcular dichas características de forma correcta. En el caso de la bradiquinesia, el procesado fue diferente, empleando una aproximación no lineal basada en algoritmos de aprendizaje automático. Después, una meta análisis fue propuesta de cara a mejorar los resultados de los clasificadores. Los resultados muestran que para la evaluación de la bradiquinesia las señales provenientes de los sensores en las muñecas ofrecían los mejores resultados. Concretamente, los clasificadores lograron hasta un 86.4% de instancias correctamente clasificadas. Analizando las señales cuando los sujetos está caminando la precisión es un poco inferior, fluctuando desde un 70.83% hasta un 75%. En ambos casos la combinación de rms+rango fue la que mostró mejores resultados. El meta análisis propuesto logró un 74.4% de acierto tomando como referencia la examinación clínica llevada a cabo por un neurólogo, fluctuando este valor desde un 91.9% hasta un 48.8%. En el estudio de la marcha, la entropía muestreada (SampEn) fue evaluada como una métrica que permitiese caracterizar la degradación de las señales de movimiento. Para ello se compararon muestras recogidas en sujetos sanos y pacientes de la EP, los valores de SampEn mostraron ser significativamente diferentes, especialmente en el caso de las señales provenientes del sensor situado en el cinturón (p<0.01) y en los brazos (p