Quantitative Research

Literature review

Android app development

Prep4Step

A personalized and secure fitness application

Timeline

24 weeks (Aug 2022 - Jan 2023)

Team

UX Project Manager, UX Designer, 2 Developers

My role

UX Project Manager

Tools

Figma, Android (Java)

In a nutshell

57%

Indians do not satisfy the WHO fitness standard

User Survey insights

I want to get fit but how can I ensure my data is safe?

Theoretical research

Theoretical research

8

8

55

55

brands

research articles

Federated learning

A secure means of providing personalized recommendations

Machine Learning

Machine Learning

A secure means of providing personalized recommendations

Flutter

Flutter

Flutter

Flutter

Android (Java)

Android (Java)

Experimenting our way through development

Wow, we made something novel and patent-worthy!

Wow, we made something novel and patent-worthy!

Impact

157

participants surveyed

Patent published

on IP India website

93.4%

step count accuracy

82.3%

recommendation accuracy

Background

Fitness standards remained unmet despite the rise in smartwatches and apps.

We started researching different domains and selected healthcare to pursue for further exploration. On digging deeper, we came across certain facts that include:

23.8 million

wearable devices were produced in third quarter of 2021 in India.

23.8 million

wearable devices were produced in third quarter of 2021 in India.

57%

Indians do not satisfy the World Health Organization's fitness standard.

57%

Indians do not satisfy the World Health Organization's fitness standard.

So, we decided to pursue fitness recommendation systems and how we could improve them.

Research

Investigated lack of user satisfaction through user research and competitive analysis.

The two statistics represented two opposite sides of the current market products. It was evident from the statistics that despite multiple options, users are still not satisfied with the fitness solutions.

How might we

design a fitness recommendation system that ensures user satisfaction?

How might we

design a fitness recommendation system that ensures user satisfaction?

User research

Surveyed 157 participants on apps and smartwatches to identify key user pain points.

To dig deeper into the issue, we conducted a survey to understand the current apps/smartwatches the users were using and what pain points they had.

On circulating a Google Form among friends and family, we received a response from 157 participants and were able to identify some valuable insights.

Key insights

Users struggled to achieve goals and felt demotivated, with many reluctant to share data due to security concerns.
Users struggled to achieve goals and felt demotivated, with many reluctant to share data due to security concerns.
  • About 61% participants used fitness bands/smartwatches and about 47% of them used it everyday all the time.

  • Average step count of 52% users was 6000 but the goal was 10000, which was not being achieved.

  • Participants were concerned about their data being misused.

  • Participants not using the smartwatches stated that accurate and good quality watches were expensive.

Competitive analysis

Analyzed 8 competitive brands used by our participants to understand their pros and cons.

On studying the brands and companies being used by our participant group, we conducted a thorough analysis of 8 competitive brands and tried to understand the pros and cons of each.

We chose the highest end model of each brand to ensure the comparison was fair. The following table shows the comparison that highlights the difference between different models.

Defining the problem

Identified the primary issues as static goals and lack of security in current apps.

With the user and market review, we understood the pain points of the users and after brainstorming, came up with the following objectives:

  • Need for a personalized system with dynamic goals

  • A secure way of analyzing the user’s data

  • Provide achievable recommendations

Problem statement

We aimed to design a secure and personalized fitness recommendation system that would provide dynamic and achievable goals to the user according to their previous performance and preferences. So, we updated our "how might we" statement to:

How might we

design a fitness recommendation system that is secure and and provides dynamic and achievable goals?

How might we

design a fitness recommendation system that is secure and and provides dynamic and achievable goals?

Literature review

Studied 55 research papers to address both personalization and privacy preservation.

The problem had two aspects: personalization and privacy preservation. A total of 55 research papers were studied by the team to cover these aspects and conduct a thorough study to understand the best way.

Fitness trackers

Reviewed 14 papers on fitness trackers, noting high accuracy in devices tested on small user groups.

Reviewed 14 papers on fitness trackers, noting high accuracy in devices tested on small user groups.

The accuracy and usage of different fitness trackers (not limited to smartwatches) was studied through a total of fourteen papers. Some devices very highly accurate but were only tested on a smaller user group.

Recommendation algorithms

Examined 9 papers on 4 recommendation algorithms to identify the most suitable one for our use case.

Examined 9 papers on 4 recommendation algorithms to identify the most suitable one for our use case.

We studied nine research papers focusing on four main recommendation algorithms to determine which is the most suitable algorithm for our use case.

Privacy preservation techniques

Explored privacy concerns by studying 3 privacy preservation techniques

Explored privacy concerns by studying 3 privacy preservation techniques

Privacy invasion was a primary concern of users and we read research papers about certain techniques that were being used and that could be used to ensure security. We studied security algorithms like homomorphic encryption, differential privacy and federated learning (FedAVG, FedSGD, etc.).

Federated learning approach

Studied 15 papers focusing on federated learning

Studied 15 papers focusing on federated learning

On comparing the different privacy preserving techniques, we chose federated learning (FL) as it was the most accurate and suitable technique.

We studied a total of fifteen papers about different types of FL algorithms to choose the best fit.

What is federated learning?

Federated learning is a privacy-preserving technique that trains models locally on users' devices, sending only insights to a remote server to enhance the global model's accuracy.

What is federated learning?

Federated learning is a privacy-preserving technique that trains models locally on users' devices, sending only insights to a remote server to enhance the global model's accuracy.

Datasets

Reviewed 10 fitness datasets and Google Fit data to understand required features.

Reviewed 10 fitness datasets and Google Fit data to understand required features.

To understand the type of data required by these algorithms, we conducted a review of fitness datasets. Features and characteristics of each dataset was studied. Additionally, we also extracted data from our own Google Fit to study the features we could use.

Design

Wireframing

Developed wireframes for core app functionality.

On identifying the requirements and technology to be used, I designed the wireframes for the required screens.

Visual design

Created a design system based on color theory and design guidelines.

Blue color was chosen as a theme as it reflects trust, calmness, stability and safety.

Quicksand was chosen as the primary font of the application as it is clean, readable, versatile and can contribute to a great user experience.

Along with the UX designer, I designed the logo following a few iterations and finalized the logo that was visually appealing and conveyed the purpose of the app well.

Prototyping

Prototyped the high fidelity designs to make them interactive.

The wireframes were then designed into interactive prototypes that would be used by the developers to design the front end of the application. I designed and prototyped the navbar to give it a flowy appeal.

Development

Developed the frontend, backend and server side of the application.

I also participated in the development process. The frontend and backend were developed using Java and the FL server was developed using Python.

Frontend development

From Flutter to Java: Adapting to Functionality Needs
From Flutter to Java: Adapting to Functionality Needs

The team first decided to use Flutter as the frontend framework. However, the available libraries did not support the functionality requirements and we shifted to Java for the Android app.

I assisted in the designing of front end elements and components.

Backend development

The backend was responsible for processing the user’s fitness data and run a machine learning model to provide dynamic recommendations to the user. It was developed in Java.

Federated learning server

The server was responsible for providing the initial training model to the user and collect insights from the users’ models and improve the accuracy of the global model.

Final deliverable

On completing the development, an Android app was ready to present.

We pilot tested the app on data of 6 people. The app was providing real-time and achievable recommendations for each user.

What next?

The concept is further being developed by an in-house team at the University to polish the product's look and efficiency further.

What next?

The concept is further being developed by an in-house team at the University to polish the product's look and efficiency further.

Impact

157

participants

157

participants

157

participants

Patent published

on IP India website

Patent published

on IP India website

Patent published

on IP India website

93.4%

step count accuracy

93.4%

step count accuracy

93.4%

step count accuracy

82.3%

recommendation accuracy

82.3%

recommendation accuracy

82.3%

recommendation accuracy

  • After testing the algorithm with data of 6 users of different physical activity levels, the algorithm achieved an accuracy of 82.3% while recommending steps.

  • The idea is patented and is published on the Intellectual Property India website.

Learnings and outcomes

  • Collaboration drives success!
    Despite challenges faced in the development, the team worked together to make sure the development was done in time. The team had to switch to a platform we were new to, but quickly learned our way through it.

  • Curiosity-Driven Innovation with Real Impact
    The knack for making something groundbreaking led us to ideate something novel and design something that would have an impact on real people.


  • Testing FL framework is a challenge
    Testing the app was a challenge due to the unavailability of data. The accuracy of the model could be increased if we had more data for testing our models.

To summarize

Designed and developed a patented app to motivate users to achieve their fitness goals!

Wearable and mobile health devices have been increasingly popular in recent years as a means of encouraging physical activity. However, the users seem to lose interest over time and are unable to achieve their fitness goals. We conducted a thorough research including reading 50+ research articles, analyzing 8 fitness brands, conducting surveys with 150+ participants. We designed and developed an application that addressed this issue by using federated learning.