Revolutionizing Healthcare with AI Assistants: A Guide to Crafting Your Own

Hey AI Enthusiasts!

I hope you're all doing well and have managed to dodge the autumn sickness season. Unfortunately, I wasn't so lucky, but after a few days off, I'm back among the living.

This edition of Bits & Bytes is particularly close to my heart, as I'll be discussing topics related to healthcare and diabetes—as well as my passion for parkour—all in the context of AI and its astonishing capabilities.?

As always, I've prepared a mix of examples based on DLabs.AI projects, as well as news from around the globe, including one of the most important AI events in our history.

Let's get started!

Introducing SugarAssist: The 24/7 Diabetes Assistant

Let’s begin with something close to home. In a previous newsletter, I shared details about a student agent we developed based on GPT technology. Now, true to our innovative spirit, we’ve turned our attention to crafting solutions for people with diabetes.

For those unfamiliar with my personal journey, here's a brief recap: Diagnosed with type 1 diabetes at the age of 10, I discovered the profound impact of data, which led me to establish a company focused on developing Suguard, an application for diabetes management. And to this day, DLabs.AI remains dedicated to enhancing the lives of individuals with diabetes.

Although our work often involves projects across various industries, our commitment to diabetes-related initiatives remains steadfast. This year, in collaboration with a team of physicians and business consultants, we’ve crafted a comprehensive Diabetes Management Guide called SugarScout, tailored for patients and their families.?

This guide is particularly helpful for parents and children navigating the challenges of a diabetes diagnosis for the first time, providing essential knowledge and support to manage the emotional and practical aspects of life with diabetes. We encourage you to download this resource if it could be beneficial to you or someone you know.

Now, why am I mentioning all this? The advent of advanced language models has opened up new possibilities. Leveraging the extensive research compiled for our eBook, we envisioned an interactive chatbot that could instantly field questions about diabetes, drawing from our validated content. Moreover, this innovation addresses a critical obstacle we’ve faced: the scarcity of comprehensive data needed to develop an all-encompassing diabetes management tool.

Hence, we created SugarAssist—a 24/7 diabetes assistant chatbot. Interested in how we brought this idea to life? I hope so because you can find the key steps of the development process below:

Creating SugarAssist: Step-by-Step Process

1. Preparing Product Requirements: We initiated the development process by hosting an event storming workshop to create a comprehensive product requirement document that laid the groundwork for our project.
2. Tool Selection: After considering our development needs, we selected Flutter Flow as our low-code platform of choice, allowing us to build a fully functioning app. As for large language models, we chose GPT-3, GPT-3.5, and GPT-4—due to their superior ability to understand prompts and generate text, which was crucial for our real-time chat assistant.
3. Defining Input Prompts: We used our eBook, Sugarscout, to provide the model with verified knowledge, converting key content into input prompts that would guide the assistant's responses.
4. Generating Responses: We developed specific rules and guidelines to direct the model on how to formulate responses (especially critical in the context of providing medical advice).
5. Retraining the Model: After initial setbacks with prompt and rule-based responses, we opted to refine the model using embeddings, which promised a deeper understanding and more personalized interactions with users, especially concerning the multifaceted nature of type I diabetes.
6. Testing Applications with Real Users: Thorough bug-bash sessions assessed the quality of interactions, identified errors, and evaluated the assistant's performance under real-world conditions.
7. Fixing Bugs and Initial Release: Post-testing, we addressed any issues and released the initial version to a targeted user group consisting of individuals with type I diabetes, collecting their feedback for further refinements.
8. Publishing the App Online: We published SugarAssist online, committing to continuous testing and enhancement to better serve children with diabetes and their caregivers, all while collecting valuable data for future support initiatives.

As I mentioned above, although the application is already available, it’s still in the testing and development phase, so please feel free to test it and share your feedback :)?

What’s most interesting about this project is that we used the power of LLMs not only to create a solution based on this model but also to work on the project itself. Among other things, GPT helped us generate a custom action in the Dart language.?

If you’re interested in learning more about this (as well as the details of the project), read our article on How To Use LLMs in a Software Developer’s Work [Case Study Included].

By the way, recent news from OpenAI shows that creating assistants to address the straightforward needs of your internal team may soon become a trivial task, thanks to new ChatGPT features that enable the creation of customized GPTs.

However, in our case, I'm talking about an advanced system that can provide detailed, accurate responses to queries from your database. It's designed to minimize the production of hallucinatory content and is suitable for commercial use.

I’m going to publish a series of videos to show how to use these features and demonstrate their business opportunities, so stay tuned!

Voice Recognition as a Diagnostic Tool? A Surprising Possibility of Detecting Diabetes

We're sticking to the topic of diabetes, but this time, let's shift our focus from our initiatives to a pioneering study by Klick Labs.

Have you ever envisioned a world where speaking into your smartphone could reveal whether you have diabetes? This is no longer the stuff of science fiction; it's the substance of a profound study that intertwines voice technology with artificial intelligence.

Klick Labs researchers developed a smart program that can tell if someone has Type 2 diabetes by listening to them talk for just a few seconds. They combined short voice clips with basic information like age and weight to create a tool to spot the condition.?

And the result? 89% accuracy for women and 86% for men. To conduct the research, they asked over 250 people, some healthy and some with Type 2 diabetes, to say phrases into their phones six times a day for a couple of weeks. With over 18,000 recordings, the researchers examined various sounds in the voice that are different for people with diabetes.?

The researchers didn't just listen to the way people talk. They used technology to find subtle changes in tone and loudness—changes we can't hear ourselves—that can be clues to diabetes. Moreover, they noticed these changes are not the same for men and women.

Normally, checking for diabetes means a trip to the doctor and some blood tests. But this voice tech could make it possible to check for diabetes anywhere and anytime. With such promising results, the team at Klick Labs wants to keep testing this approach and even try it for other health issues.?

I’m curious what similar research would look like for this disease. I'm keeping my fingers crossed for the success of this project, and I'll be sure to keep you updated if I learn anything new.

Source: Mayo Clinic

Digital Transformation Success Story: How CMH Overcame the Digital Disconnect

And now, the topic of assistants again, but in a broader context. I'd like to share with you an amazing case from the healthcare sector where Albany Med-Columbia Memorial Health (CMH) is redefining its approach to healthcare delivery.

CMH confronted a challenge that many healthcare providers fear: the digital disconnect. Despite serving over 100,000 residents, their digital shadow failed to capture the entirety of their market share. Patients, symptomatic and in search of care, were turning to Google, and they were shepherded towards other healthcare systems with more robust online presences.

CMH's response? A revolutionary AI-powered care navigation assistant, accessible round-the-clock on their digital platforms. By deploying it, CMH streamlined patient triage, enhanced the care journey, and, notably, boosted its business outcomes.

Now, let's talk numbers: With the system live, CMH has reported a threefold return on investment. Patient feedback has been overwhelmingly positive, averaging a Net Promoter Score of 72. In 12 months alone, they've seen over 9,000 completed assessments, with nearly 40% of patients returning to the system for their healthcare needs.

This initiative is not just about improving the quality of care but also about strategically positioning CMH in a fiercely competitive market. The organization's marketing efforts, blending SEO and SEM strategies with traditional offline promotion, exemplify a holistic approach to raising awareness about their innovative care navigation assistant.

From my vantage point, CMH stands as a testament to the transformative power of AI and highlights the growing influence of AI across industries. As I engage with our clientele, a common thread emerges — a concern over being left behind in the digital race. Virtual assistants represent a beacon of opportunity, extending well beyond healthcare.

Should you find yourselves on the cusp of such an initiative and need guidance to take that leap — I’m here to help. Just leave a comment or drop me a message. Remember that in this era, only the bold survive.

Source: Healthcare IT News

Google Unveils AI-Powered Search to Simplify Medical Record Retrieval for Clinicians

As for the application of AI in healthcare, Google is diligently forging ahead. Their latest development from Google Cloud is an AI-powered search tool poised to streamline healthcare workflows.?

This innovation offers clinicians a seamless way to extract clinical data from a myriad of sources including electronic health records, clinical notes, and even scanned documents into a single, accessible location. The potential benefits are significant: reduced administrative burdens for healthcare providers, who often face the daunting task of combing through various systems and formats to find critical patient information.?

By simplifying the search process, Google Cloud's tool aims to alleviate the frustration and time drain associated with clinical information retrieval, freeing up healthcare professionals to focus more on patient care. The creators highlight that the tool is designed to save time, reduce clinician frustration, and provide easy access to accurate information — all while ensuring data accuracy with direct links to original sources.

Google's new search tool will be integrated into the Vertex AI Search platform, extending Google's existing suite of healthcare products. Early testing by healthcare giants like Mayo Clinic demonstrates a cautious yet optimistic approach to adopting AI in patient care, emphasizing the importance of a tool that fits seamlessly into existing workflows without adding friction.

With HIPAA compliance and a no-data-sharing policy, Google Cloud ensures the privacy and security of patient information. This technological leap could not only streamline healthcare operations but also enhance the overall patient experience by connecting the dots in a historically fragmented system.

This solution is about real-world impact, about slashing the bureaucratic red tape that has led to an alarming rise in physician burnout. With Google's search capabilities, the burden of paperwork could be substantially alleviated, giving healthcare providers what they crave most: time.

Time that can be redirected towards patient care, time that can heal rather than hinder.

Souce: CNBC

Trustworthy AI: OpenAI's Strides in Making LLMs Understandable

Curious about what's going on with the creators of ChatGPT? Well, if you've been following my newsletter for a while, you may be familiar with topics around explainable artificial intelligence. Interestingly, OpenAI is now making progress in that very direction.

They’re developing a tool that will deconstruct the mysterious decision-making process of large language models (LLMs) like GPT. The tool aims to identify which specific parts of an LLM are influencing particular behaviors. Though still in its infancy, the tool's code has been made available open source on GitHub, marking a significant step towards transparency in AI.

The process involves isolating the "neurons" of a model, akin to identifying patterns in the brain, to determine their role in generating responses. OpenAI's tool dissects the model neuron by neuron, using the company's GPT-4 model to explain the actions of these neurons and to evaluate the accuracy of these explanations.

By applying this approach, the team at OpenAI has managed to generate natural language explanations for the function of each neuron and provide a reliability score for these explanations. This pioneering research has yielded explanations for all 307,200 neurons in the GPT-2 model, compiled into a comprehensive dataset alongside the tool's code.

While the tool is in early development and confident explanations cover only a fraction of the neurons, it holds promise for future applications such as reducing biases and enhancing the performance of LLMs.?

Jeff Wu and his team acknowledge that while many neurons still elude a clear pattern or explanation, this tool could evolve to tackle not just the interpretability of static models but also those, like GPT-4, that harness the internet to enhance their capabilities.

Well, I hope this research will pave the way for a future where AI decisions are as clear and trustworthy as possible.?

Source: OpenAI

Big Meeting on AI Safety at Bletchley Park

I couldn’t fail to mention the historic meeting on AI. In Bletchley Park, where the Enigma code was once cracked, the world's first AI Safety Summit unfolds, drawing representatives from 27 nations, including the U.S. and China, as well as industry heavyweights like Elon Musk and OpenAI’s Sam Altman.

Under the auspices of Prime Minister Rishi Sunak and the U.K. government, the summit marks a pinnacle in the escalating global discourse on AI safety, sparked nearly a year ago by the launch of ChatGPT. While the summit stopped short of pressing for a consensus on enforceable AI safeguards, it made significant strides.?

PM Sunak heralded an agreement for early government access to AI models for safety evaluations and appointed computer scientist Yoshua Bengio to spearhead a consensus on AI system risks. Despite the absence of granular details in Sunak’s announcement regarding the nature of this early access, the discussions set the stage for the inception of the U.K.'s AI Safety Institute, signaling a lasting commitment to AI evaluation and governance.

The event, symbolically initiated with the "Bletchley Declaration" on AI, saw the U.S. assert its stance on AI safety. Vice President Kamala Harris outlined American initiatives that, in some respects, seemed to surpass the U.K.'s efforts, highlighting a complex dynamic of collaboration and leadership (when it comes to U.S., it's worthy to mention that President Biden has issued an Executive Order aimed at ensuring the development of safe, secure, and trustworthy Artificial Intelligence).

The summit also reignited the enduring debate over open-source versus closed-source AI research, dividing the community over the best path to safer AI development. Interestingly, at the heart of Bletchley Park, as security personnel kept watch, a small group of activists from Pause AI made their dissent known, calling for a stop to AI advancements beyond the capabilities of GPT-4.

Although no major decisions were reached, I am pleased to see that AI is being taken increasingly seriously and is receiving considerable global attention. Regardless of whether it is as insecure as Elon Musk has once again emphasized, regulation in this arena is surely necessary to ensure our collective safety.

Source: Time

Navigating Obstacles: How AI-Powered Robots are Mastering Parkour

Finally, I wanted to share a small project related to my area of passion - parkour. The incredible agility and coordination seen in human parkour have long been a source of inspiration for robotics.?

Achieving this kind of dynamic movement in robots typically involves the precise engineering of separate systems for perception, actuation, and control. This often limits their operation to highly controlled environments, such as lab-based obstacle courses with predefined layouts. Humans, on the other hand, learn parkour through practice, adapting their innate abilities rather than re-engineering their biology.

In a fascinating development, researchers have embraced a human-like approach to teach a small, cost-effective robot parkour. Equipped with basic actuation and a rudimentary depth camera, which only provides low-frequency, imperfect visual data, the robot faced significant challenges.?

However, by employing a single neural network policy trained through large-scale reinforcement learning in simulation, the robot learned to navigate a course with remarkable precision, directly translating the imperfect sensory inputs into refined control outputs.

This robot demonstrated not only the ability to clear high jumps over obstacles twice its height and long jumps across gaps twice its length but also performed complex maneuvers such as handstands and navigating slanted ramps.?

Impressively, it could also adapt to novel parkour courses with varying physical properties.

A case study published on GitHub delves into extreme scenarios, showcasing the robot's prowess in a parkour course and more. The practical applications of such agile robots are vast, particularly in scenarios that require access to hard-to-reach areas.?

Whether it's navigating the rubble in search and rescue missions, performing maintenance in intricate industrial settings, or exploring uneven terrain in scientific research, the implications of these advancements in robotics are as exciting as they are vast.

Source: GitHub

___

I hope you enjoyed this newsletter, even if you’re not a fan of healthcare ??

I'd also love to know which topics interest you the most, so please let me know in the comments.

Stay healthy, and I'll see you next month.