1 Introduction

Parkinson's Disease (PD in short) is a special kind of neurological disorder that progressively impairs the ability to control movement. It is characterized by tremors; rigidity of the muscles; postural instability; slowness of movement; stern facial expression and speech; and difficulty with balance and coordination [Sveinbjornsdottir, 2016]. With the progress of this disease, these symptoms usually get severe and may interfere with the ability to perform regular daily activities [Shulman et al., 2008]. The primary cause of Parkinson's disease is yet to be identified. However, both genetic and environmental factors are considered responsible for Parkinson's disease [Morale et al., 2008], [Warner and Schapira, 2003]. In this project, we target the people who have recently been diagnosed with PD and introduce a virtual avatar that can help them out with their queries and support.

1.1 Problem Statement

As we have already discussed, people with Parkinson's disease usually suffer from limited motor functionalities. As a result, they find it difficult to type or use a computer and look for resources about Parkinson's Disease which are not well compiled. However, when diagnosed with PD, an individual gets overwhelmed with many questions regarding the disease – what causes PD, what are the symptoms, how PD is treated, how it will affect other family members, etc. At that time, they feel the urgency to get some instant resources, preferably from a human expert. However, there has been a significant scarcity of neurologists, even in developed countries like the United States. According to [Dall et al., 2013], the average wait time for an appointment with a neurologist is 34.8 days. Therefore, we can think of a better solution for these newly diagnosed PD patients – provide them the opportunity to ask their questions to a virtual avatar that has access to a variety of vetted resources about Parkinson's Diseases. Figure 1 is showing a user interacting with the avatar to get an insight about Parkinson’s Disease.

1.2 Motivation

In general, we may deem it natural that an individual will always feel comfortable reaching a human being for their queries regarding any unknown matter. However, research shows that PD patients have reduced trust in human social interactions. According to a study conducted by [Javor et al., 2016], PD patients had higher faith when speaking with an avatar than when talking with a human regarding the same subject. There are a few potential reasons for this. First, avatars are non-judgmental listeners that can provide undivided attention. Second, avatars can be customized to look and sound like the patient's ideal therapist, which may make the patient feel more comfortable. Finally, avatars may be able to provide more personalized and tailored treatment than human therapists. We hope to be part of the solution in overcoming behavioral deficits and offering PD patients a solution that accommodates their neurobiological needs.

The significance of a virtual avatar can be comprehended from the perspective of post-diagnosis care and follow-up of Parkinson's disease. To keep the disease under control, the patients must ensure that they take their medication as prescribed. Since PD is a progressively deteriorating disease, alteration in medication may introduce increases and new symptoms. There are also many side effects of the medication that the patient may experience, so it is crucial to monitor them closely. Other medical implications include ensuring the patient is getting enough exercise, eating a healthy diet, and getting enough rest. A dedicated virtual avatar can be a personal caregiver and keep track of a PD patient's medication, symptoms, and daily routines.

One of the major advantages of a virtual avatar is that it can provide information about Parkinson’s disease to individuals who are friends or family of a PD patient. So that people who find it really difficult to visit a neurologist in-person, perhaps for the long appointment delay, for being physically not capable of going to a doctor, or even due to economic conditions, can easily get a regular follow up by just using the virtual avatar.

1.3 Current state-of-the-art

The current state-of-the-art is browsing the web and retrieving information about Parkinson's disease manually. However, the resources are minimal. A bunch of information is available online, but it can be challenging to find reliable sources. It is also essential to be careful when surfing online media, as there is a significant amount of misinformation. If anyone is looking for reliable information about Parkinson's disease, the best place to start is the Parkinson's Foundation (https://www.parkinson.org/resources-support) website. They have a lot of resources and information about the PD condition, treatment options, and support. However, the potential issue with this website is that it can not provide information in real time. Individuals, albeit seeking urgent needs, must send an email asking their queries and then wait for the response.

Our goal with NeedFinding is to find product opportunities for our virtual avatar for Parkinson’s patients. Coming from a user centered approach, we wish to understand our users’ expressed and latent needs.

Participants representative of the intended user base will be recruited through team member Masum's lab, as he has an existing network of Parkinson's patients. Domain experts, ideally neurologists or primary care providers, will be recruited through team member Michelle's medical school program.

Once we map out how the communication will work and how our prototype will communicate based on its use cases we plan to iterate through a process of prototyping, testing and redesigning. Since the “frontend” of our project has already been completed our prototypes will focus mostly on the backend of our project that being how our prototype communicate with the user, some of the major concerns that will be addressed in these prototypes are how we questions will be recognized and answered but also how our prototype will simulate talking with an actual person which includes things such as positive reassurance and generally more human like responses that would make a patient more comfortable given the fact that they are chronically ill with Parkinson’s disease.

To create such prototypes, we plan to use techniques such as storyboarding and wizard of oz to specifically demonstrate and receive feedback on how the bot communicates with patients with Parkinson’s disease. This process of receiving feedback and revamping our prototype by seeking input from professionals and individuals with Parkinson’s will allow us to tackle obvious problems early on and will increase our chances of success dramatically later on. These prototypes of the backend features will start off low fidelity and will increase to higher fidelity while any depiction of the front end will be high fidelity as that part is already done.

As mentioned, because user feedback on accessibility is critical, we’ll want to have some semi-structured interviews with each round of participants. We want to be cognizant of the Hawthorne effect, as well as effects of repetition on the user’s proficiency on our technology.

Thusly, for every given iteration of prototyping, we’ll interview 5 people who fit our user personas. That is, we want 5 individuals who are already presenting with features of Alzheimer’s or other neurodegenerative diseases at any given round.

However, because of the small populations of people who are presenting Alzheimer’s symptoms who are able or willing to provide informed consent, we believe we’ll have to be judicious, and balance our desires for measuring a true baseline of accessibility against maximizing our population size.

So, group design for our feedback will be constructed thusly:

1. We’ll have 3 rounds, based on the 3 levels of prototyping that we’re trying to implement.

2. 5 people randomly selected that fit our persona and established user. From here, 4 people will be selected based on a criteria to be part of a repeated group, while 1 person will drop out

3. 4 people from the prior round, 1 person drop out, 1 person enter prior to new interaction. We will bias the weighting of everyone’s opinion such that the 4 people’s opinions are 60% of the determination, and the 1 person is 40% of the last determination.

4.1 Technology stack

In our implementation, we are using a natural language pretrained model which will semantically match with the user's query from an existing database. There are multiple ways to do this, and the methods are not well-established in the literature. Hence, we will have to experiment on our own to find the best method for our particular implementation. Few ways to do this are,

1. Sentence encoders

2. Pretrained Natural Language Inference models

3. Long-form question answering system

The NLP model will fetch the appropriate question answer pair from the database, and the virtual avatar will speak it out.

4.3 List of Features

A user can ask any question about Parkinson’s disease and a virtual avatar will answer them. The user can speak directly in natural language just like a human conversation. The avatar can make reasoning on multiple answers in the database and answer what the user requires. It can provide emotional support to the user if needed.

4.4 Technical Challenges

Semantic searching is an active area of research in NLP, and not fully solved. Matching the user’s query with an existing database might not always work due to the nature of semantic search models. The virtual avatar and speech recognition and generation systems are outside API’s which might fail.

4.5 Identify Possible Solutions

In case of failure of the semantic matching language models, we would do a plain text matching instead. In case the virtual avatar or the speech system fails, we will make it into a text based system, where you would interact with the database of questions related to Parkinson’s through typing on the keyboard.

4.6 Backend

The backend of our system will be a database of question-answer pairs related to Parkinson’s Disease. An NLP model that matches a user query with the database and fetches the right answer. This answer is then sent to the virtual avatar for delivery to the user. The virtual avatar here works as the front end, and everything else is back end.

4.7 Platform

Our system will be available to any device with a video calling feature. Primarily we are implementing it over zoom, as it is one of the most popular video calling platforms out there. Making this into an on-device app would limit its accessibility and introduce many device dependencies.

5.1 Evaluation Hypotheses

Our evaluation hypotheses are:

1. A virtual avatar can provide a more human-like experience with things such as positive reasurment that surfing on the web cannot.

2. Interacting with a virtual avatar for Parkinson’s disease related information is more convenient and user friendly than browsing the web.

3. Interacting with a virtual avatar can provide more specific information than looking for it online.

4. Interacting with a virtual avatar involves less physical activities. As a result, for an individual with Parkinson’s Disease, it is more easier to use than actively using computers or other electronic devices to search for information online.

5. A virtual avatar can provide information in real-time while online resources are not capable of providing instant support.

5.2 Control Condition and Participants

The control condition will be the current state-of-the-art of finding information regarding Parkinson’s Disease that means searching from the internet sources. We will recruit patients with Parkinson’s disease. Since our prototype’s purpose is to help those with Parkinson’s disease it is only natural to recruit those with parkinsons. By running experiments with intended users we minimize the risk of failure. We plan to reach out to URMC through the ROC-HCI lab. We plan to connect with some Parkinson's Disease patients and ask them to voluntarily get involved in our evaluation process. We want to get as many participants as possible.

5.3 Qualitative Experiments and Evaluation Metrics

We plan to conduct the within-subject experiment to evaluate our proposed system against the state-of-the-art. We intend to recruit real users, including individuals with PD and without it, to complete the experiment. As we have already discussed, the current state-of-the-art of finding information about Parkinson's Disease is to surf the web and look for an authentic source of information. Therefore, our participants will be asked to gather their required information from the web as well as from our virtual avatar. Since there is a chance of introducing bias based on the ordering, the web-based and virtual avatar systems are present to them. Therefore, we will split the participants into two separate groups and change the order of the tasks to minimize bias. As evaluation metrics, we plan to collect the user satisfaction level, the amount of information they can gather from each system, the authenticity of the gathered information, etc.

5.4 Risk Management Plan and Fallback Plan

Our proposed solution not panning out will mean that we fail to establish that a virtual avatar is a better platform for disseminating information about Parkinson’s disease. As previous studies [Javor et al., 2016] have established otherwise, it will still be an interesting finding.

Alternative to our proposed system, if some implementation fails, we will make it a text based question answering system. In case of failure of the semantic matching language models, we would do a plain text matching instead. In case the virtual avatar or the speech system fails, we will make it into a text based system, where you would interact with the database of questions related to Parkinson’s through typing on the keyboard.