Study on what goes into making an AI-based product
AI-based Web App for Workplace Mental Health Prediction and Intervention
Primary role:
AI Product Designer
Time frame:
2024 (3 months)
Industry:
AI and Healthcare
Responsibilities:
End-to-end process from intial research to project proposal
Project Overview
This project focuses on the development of an AI-driven platform that monitors employee mental health by collecting biometric data from wearable devices like Fitbit, Apple Watch, and Oura Rings. Leveraging a Deep Neural Network (DNN) model, the platform analyzes data such as heart rate variability and working hours to predict mental resilience and identify employees at risk of stress-related issues.
This project was undertaken as part of my capstone research at the MIT-AI program. Upon submission, it was recognized by my instructors and teaching assistants with an exemplary work certification, highlighting its innovative application of AI and potential for significant impact on workplace well-being.
Overview
Mental health issues rising among corporate employees
Mental health issues among corporate employees, including stress, depression, anxiety, exhaustion, hopelessness, and burnout, have been on the rise, and have recently exacerbated by market dynamics such as layoffs.
According to Forbes report 2024:
74%
Employees report negative mental health at work
1/3
Americans say work is adversely impacting their mental health
80%
Feel stress at work and unable to fuction at their usual level
Employers are worried about decreased productivity, increased absenteeism, lower retention rates, increased costs, and a negative impact on morale.
Employers need a platform to support employees in addressing mental health challenges and enhancing their overall well-being
Building an AI-based product to support employees
With the current advancements in the Artificial Intelligence (AI) technology, we designed a product that can make the best use of it to support employees' health. AI provides a novel way to analyze the different symptoms indicating poor health of employees, so that employers can timely provide them the required support. However, before finalizing the product, it's essential to understand what it means to build an AI-based solution and how it will serve end users effectively.
Defining scope: health management system for employees using biometrics data
Biometrics data collected from wearable devices (such as Fitbit, Apple Watch, Oura Rings, etc.) have proven to provide good indications of health issues. For our product, we collect user-approved biometric data and analyze those to provide the best health recommendations. Our intelligent solution focuses on tracking data collected during working hours, and provide the optimal signal towards employees' mental and heart health (like heart rate variability).
As part of our health support measures, we offer potential solutions to users who are currently affected or are at higher risk of health issues. The product dashboard includes quick at-work options like 1-minute workout sessions and breathing exercises for calmness, or long-term psychological assistance by connecting them with respective employees.
Objective metrics: measure the effectiveness of the solution
Intelligent solutions need to be measured to define their success criteria. For our product, we identified the following computable metrics, which will be continuously tracked to see its effectiveness.
Accuracy
Accuracy measures how well the solution predicts mental health outcomes, such as mental health resilience or identifying potential health issues
Precision
Measure the precision of the solution in correctly predicting positive cases, such as how many employees truly suffer from stress, anxiety, or any other mental health issue
Specificity
Assess the solution's ability to correctly identify employees without mental health concerns or identify those specific cases where they suffer from any specific mental health related issue.
Strategy: Comprehensive customer solution package with integrated AI capabilities
Our product is targeted as a comprehensive solution package for corporations, offering a full range of health-related services to support employee well-being. Powered by AI, our product is one of the best providers of these services.
We use AI as the driving technology to enable providing those services.Our produt integrated features like video recommendations of exercises to improve their mental health, 1-minute workout plan that users can perform from their workplace, connecting them with clinicians who need extra attention, and providing a special health insurance plan to those who need extra help.
All of this would be founded on the AI technology that derives meaning from health data input recorded from wearable devices.
Operational requirements
Before building a product, we need to identify its operational requirements to make it feasible and successful. For our product, we need to analyze :
Budget and Cost Analysis
Timelines and Roadmap
Specialized Design and Research Team
Specialized ML Engineers and Data Scientists
Choosing an AI Techonology
With the numerous AI technological options available, we need to chose one that's best suited for our purpose. Deep neural networks (DNN) have emerged as one of the most promising fields, demonstrating high success rates across various domains. Research has shown that it outperforms other conventional AI methods like SVM and Tree-based algorithms. It also eliminates the need to create specialized architecture. Our focus will be gathering a substantial amount of data and feeding it into the DNN model for effective training.
The diagram above represents a sample deep neural network architecture, where input to the AI model is biometric data from wearables and the output are potential health signals from the model. The output signals are used for tracking the health status of employees and providing them with the required support. The model above shows three convolutional layers, which extract patterns from the input data. The convolutional layers are followed by dense layers, which compute correlation between different patterns and generate the final health signals as the output.
Strategic Data Collection: managing hardware to enhance control over user data
AI algorithms hold no meaning without an appropriate dataset that represents the problem being solved. Our data collection approach involves users wearing devices, allowing us to collect data directly from wearables available on the market. The primary challenge lies in encouraging users to wear commercially available hardware. For the user-approved scenarios, the wearables allow us to periodically collect quality data samples. Additional user metadata can be gathered during the registration flow, such as age, weight, and gender. Our data collection strategy involves encouraging employees to wear the hardware, keeping it synced in real-time with our servers, and ensuring continuous feedback.
This approach is followed by many tech companies like Apple and Google with their health monitoring wearables and the associated software applications.
Fourth Stage - Tinkering
Development Flows: AI solutions needs long-term infrastructure and experts
By this stage, we have a strong idea of how an AI-based solution can contribute to the success of our product. But AI solutions are not "built once"; they need continuous updates to accommodate the changing nature of users and regulations. The solution needs to adapt to the changing environment and be up-to-date with the latest technological advancements to remain competitive. To make an AI-based solution successful for a product, we need
Strong Infrastructure
Infrastructure supporting large scale distributed operations and hardware accelerators like GPUs
Domain Experts
Team of experts who understand the technological advancements and can quickly incorporate them into our product.
AI Cancers
AI offers a revolutionary way to solve human problems, but one needs to be careful when deploying this powerful tool. AI-based solutions come with their equally complex challenges that could have catastrophic impacts on outcomes. AI developers and product owners need to be aware of the AI cancers before making employing this into end-user products. For the health monitoring systems, we need to strongly look at the following
AI algorithms attempting to solve any given task, like detecting stress, can easily be overfitted to samples seen in training data. Development flow like dataset processing and training should be aware of these and ensure the models are trained to represent a generalized task.
AI algorithms are prone to being biased towards particular solutions. In the figures above, depending on how the data samples are "seen" by the algorithm, the resulting AI solution can vary vastly in how it processes the data. This can be very dangerous for the health monitoring systems in particular if say it fails to distinguish between male and female samples for any given type of biometric data.
Concept to Design: Blueprint for an AI-Powered Dashboard
I created a high-fidelity prototype as shown below. I designed the screen with all the micro-interactions. The clickable prototype helped the developers to understand the micro-interactions in the application during the development phase.
Product Blueprint