Hi everybody, my company is called UltraViolet and our goal is to build a platform that integrates all health information (from wearables, medical records, any tests) in order to give you personalised advice on any health concern that you might have or if you want to optimize! We want to make this process easy and enjoyable as we assist on your health journey.
If you have ever been confused by your metrics our AI tool is for you!
This is pretty neat! I uploaded a few PDFs (including a blood test) and it worked flawlessly.
I was a little hesitant to upload more because I’m unsure about your privacy policy. Also, do you incorporate the ChatGPT custom instructions? It’d be nice if I didn’t have to re-upload the same data each time I want to ask a new question.
It was interesting to try this. Honestly, I didn’t notice skill that regular ChatGPT doesn’t have, but I didn’t make a careful comparison to see what your training had added. What do you think it’s going to be best at?
In my test, UltraViolet got to the same answer I did when I originally did the analysis manually. Even though I had already done some of this exploration, I tried to interact a bit naively and leave more work for the AI. I had to coach it through the steps to get it to stop thinking generically (god these AI’s are verbose) but when I asked it to co-pilot with me it did a normalish job.
I think the process of analyzing self collected data is going to be involve a lot of robot/human teamwork, so anything that can be done to train on known problem-solving steps will help, including topics like aggregation/binning of data, characterization and simplification of records, and turning hypotheses into analysis will be useful. Thanks for sharing this resource with us!
Also, I’ve noted in many trials over the past year that ChatGPT absolutely loves to say correlation does not imply causation Fortunately, I know it does not really believe this.
You
If I’ve gained weight over the last year and want to know whether my newish habit of eating late is a cause, and I have data for all my eating times as well as daily weight, what is the simplest way to find out if this hypothesis is worth pursuing further. Please limit your answer to 200 words.
ChatGPT
The simplest approach is to perform a preliminary analysis using correlation and basic trend observation:
Calculate the average eating time each day. If you usually have multiple meals or snacks, compute the weighted average based on calorie content to get a single daily metric.
Plot this average eating time against your daily weight over the same period. Look for visual patterns; an upward trend in weight as eating times get later suggests a potential relationship.
Compute the Pearson correlation coefficient between your daily average eating time and weight. This will give you a numerical value indicating the strength and direction of the relationship.
If you find a moderate to strong positive correlation (e.g., coefficient above 0.5), this suggests that your hypothesis of late eating contributing to weight gain is worth exploring further with more detailed statistical analysis.
When I ask for some refinements; say, accounting for time lag, it makes many good suggestions, demonstrating a common sense understanding of how valuable correlation is for inferring causation.
From the google doc it seems great at suggesting interventions/causes. Checking for hourly pattern was a good catch. Compared to an automated exploratory analysis app I have imagined, this adds more domain expertise and easier input for end user but has fewer plots, simpler techniques and could be more thorough. Does it print code for user?
It mostly treats the data as iid and does not check for (from finding relations between time series) Change point or breakpoint detection. Stationarity. Repeated shapes. Lag. What if eating pizza on one day causes heartburn the next? Build up. What if it takes two days of eating pizza to cause heartburn?
Its very possible that glucose has an effect on arrythmia but it is only visible when other stronger causes (including the non-iid issues) are accounted for.
“Notably, on February 27 and 28, 2023, when the number of arrhythmia incidents was relatively low (4 and 9 incidents, respectively), the average blood glucose levels were at their highest (128.34 mg/dL and 131.54 mg/dL, respectively).” This statement is really too simple.
" * Patterns: While there’s variability, a pattern of low blood glucose levels with rising trends and high levels with falling trends during arrhythmia-prone hours could suggest that sudden changes in blood glucose levels, whether increases or decreases, might be associated with arrhythmia incidents."
Some internal mechanism in response to arrhythmia? Arrhythmia caused by return to otherwise normal functioning?
" Can you characterize the pace of change in these 20 hours " The AI misunderstood and talked about the 24 hour hours. And I suspect the later correlation discussion was marred my this issue.
