I was feeling tired a few weeks ago, with a runny nose, sore joints, even watery eyes. So, like the 70% of fellow internet users who conduct a Google search before heading to the emergency room, I did a quick search of my symptoms.
The results were as expected: maybe a common cold, maybe sinusitis, probably allergies, marginal chance of pneumonia or the flu. And of course I experienced a fleeting wave of COVID-19 panic, even post-vaccination. But what I didn’t tell Google was this highly relevant information: it was Monday after a long weekend entertaining my in-laws, I was run ragged in a Sunday soccer game by a bunch of 20-somethings, and the autumn leaves were finally changing here in Boston (yes, people, fall allergies are a thing). In short, I quickly realized I had a classic case of the winter-is-coming-and-I’m-getting-older Mondays.
If an internet search of “stuffy nose” has also informed you that you could be dying, you’re not alone. There are thousands of memes chronicling these misadventures, and more seriously, studies link repeated health-related internet searches to anxiety in a clinical phenomenon known as Cyberchondria. Even with increased access to data and knowledge, medical diagnosis using technologies such as artificial intelligence and machine learning is extremely challenging (evidenced by Google Health’s storied history, for one).
This is in large part due to the fact that these search engine diagnoses are often possibilities without associated probabilities. Sure, the possible ailments are sometimes categorized with words like “rare” or “very common,” but much more research is necessary to determine how—and how much—our symptoms predict actual onset of disease. This is where simple ‘googling’ falls short and more sophisticated AI is needed.
That’s why I’m excited about Stratyfy’s recent role in a ground-breaking, peer-reviewed study on the prediction of viral symptoms using wearable technology and AI, led by Rockefeller Neuroscience Institute (RNI) in collaboration with West Virginia University Medical Center, Vanderbilt University Medical Center, and Thomas Jefferson University. This is the first study of its kind using wearables and apps with machine learning to predict viral illness-like symptoms, and it could be a game changer for limiting transmission of viral illnesses like COVID-19 in the future.
For the study, volunteers wore a smart ring device to collect physiological measures and used a mobile health app to track self-reported symptoms, social exposure to COVID-19, and other factors such as workload and fatigue. With this data, the study’s model was 82% accurate in predicting the likelihood of developing symptoms consistent with a viral infection—three days before the individual even experienced them. As we look for ways to slow the spread of viruses like COVID-19, this finding is critical, since accurately detecting possible infection this early allows for effective quarantine before symptoms occur. The model’s framework also minimized warnings to individuals with a low likelihood of developing viral symptoms—which could mean a future where we rely less on “abundance of caution” (and in my case, googling symptoms) and more and more on data-driven decision making.
The study’s model was developed on Stratyfy’s Probabilistic Rule Engine (“PRE”), our proprietary modeling approach, selected for “the transparency and interpretability of our model,” according to the researchers. While the building blocks of PRE-based models are rules, Stratyfy uses an additional parameter to enhance these rules—which means that PRE’s predictive power is on par with black-box machine learning approaches, without sacrificing the transparency and flexibility of a rule-based model.
PRE’s flexibility was key to this study in its ability to utilize a variety of inputs and go beyond a singular source of information. The model used physiological data from the wearable devices, custom mobile health app self-reported data, AND expert medical knowledge from both professionals involved in the study and research from peer-reviewed sources outside the study. In short, using PRE we were able to not only preserve—but more importantly, leverage—the human experience, both of the patients and doctors, to build a stronger, more accurate predictive model.
Now, with the increasing overlap of fintech and digital health, our technology’s power in this work is no surprise. But our involvement in this study is more exciting than just the results. When we built Stratyfy, we set out to use PRE to tackle complex problems like these —questions that historically have been difficult to solve with other methods. After all, existing AI solutions are only as good as the data used to teach them—data that is often plagued with biases.
Traditional machine learning approaches often lack the context and insights of experts, so they cannot predict a future that differs from historical data.
As a human-first AI company, we believe that machine learning must go beyond just trusting the data, so we’ve built transparent, interpretable technology that allows companies to combine the wisdom of their people with the precision of their data.
This mission brings us to where we are today. Much like many medical diagnoses, financial decisions such as credit risk require that we leverage both the experience-driven expertise of humans and the automated, scalable power of technology. Both also require that we have insight into how decisions are being made. Nearly 1 in 4 US adults who applied for credit in 2020 were denied, and we know that we can improve these outcomes—and the financial health of millions of people—with credit risk solutions that mitigate historical bias and give lenders the tools to make more informed decisions at scale.
The findings of the RNI study not only illustrate the real impact fintech can have in healthcare (move over WebMD), but they also get the team at Stratyfy excited about the next generation of human-first AI—specifically, PRE. In any industry or use case where transparency and bias mitigation are paramount, we believe PRE is the future of decision making, and as demonstrated in this study, we’re already proving it.