I have worn my Ultrahuman ring for over a year and I’m sure if you have ever read any of my posts before, you know that checking my Ultrahuman app is the first thing that I do, even before I check my Instagram most mornings. But somewhere around month eight, I came to a realization that I have a massive amount of data, which I am doing absolutely nothing with. Sleep scores, HRV, temperature, recovery and all those resolutions of 10K steps a day, down the drain. Everything every single day, but the only insight I am pulling is “oh my score is low again, cool.”
But when my ring flagged something off with my ovulation and I proceeded to ignore it, resulting in a delayed period, I decided to actually do something about my wearable data. I exported the months of data and asked Claude, ChatGPT and Gemini to help me analyze it in order to find out what my body was actually saying about my cycle. Not some generic information that you can get anyway from the app itself. I wanted to know what my data was specifically saying, month after month and if there were some patterns that I have been too tired or too busy to notice on my own.
Here’s how that actually went, what worked, what didn’t, and what I’d tell anyone else trying to do the same thing.
The Data I Was Working With
I exported data in CSV format from the Ultrahuman app and Ultrahuman Vision dashboard for about five months. These were:
- HRV, steps and resting heart rate
- Sleep stages and sleep score
- Body temperature difference from baseline
- Readiness and activity scores
- A cycle report covering six tracked cycles, including ovulation status and luteal phase length for each
I’m not one for coding and building interactive dashboards, so I asked all three the same thing: Build me a dashboard from this and tell me what’s actually going on with my hormones, stress and recovery.
Round One: Gemini
Gemini jumped right into the code. It imported the files using pandas, performed a describe on all the columns, generated a correlation matrix and even created a four-panel matplotlib dashboard showing HRV, resting heart rate, sleep score and steps. A very good start and I appreciated that it looked for missing values first.
What followed was a great write up. It noticed the PMOS/PCOS history in my cycle report, noticed that there were three out of six cycles that had no ovulation detected and tied those factors together with insulin resistance and cortisol quite well. It gave me a figure for the HRV to resting heart rate correlation, negative point six seven and used that to explain the stress factor.
Where it lost me a little was the tone. Each section started with a header with an emoji and a bulleted list, followed by another bulleted list, another and yet another one. By the time I finished reading it felt like a newsletter than something tailored specifically for me. It relied on very generic supplement information and breathing exercises which is perfectly fine but doesn’t necessarily require a detailed analysis of my seven months of data.
Round Two: ChatGPT
ChatGPT didn’t go into much depth about the actual code but jumped right into creating the “written” dashboard including a few visual tables with current values compared to the goal for each metric, sleep score, HRV, number of steps and finally a hormone metrics table, stress and inflammation sections completely made of proxy measurements, since the ring definitely doesn’t measure inflammation directly.
To be fair, the interpretation of the data was decent. One thing that it picked up that the others were not quite able to interpret correctly was that my HRV increased drastically in comparison to earlier in the year and made it the only win in this dataset. It created a summary scorecard for the end, where I got nine out of ten for my sleep and four out of ten for daily movement, which was a nice quick gut check.
It did create a trend chart too, showing the trend of HRV against resting heart rate, but it looked like something we used to plot during the 7th grade- very basic. Finally, it asked to make a Notion-like dashboard with gauges and trend charts. While that would be fine, I expected it to create it the first time, not as a follow up offer.
Round Three: Claude
Claude was the only one that actually built the tool rather than described its capabilities. The same data, the same prompt and in response, I got a fully-fledged interactive HTML dashboard with the Card layout and the visual charts.
The data Claude extracted looked more precise as well like it noted that my HRV rose from the average value of 47 ms in December to the whopping 78 ms by June and put it in relation to my daily steps decreasing from 5,600 a day at the beginning of the year to mere 2,100 a day starting from May. I’m not proud.
Furthermore, it linked this information directly to my cycle report, mentioning that the anovulatory cycle with the atypical luteal phase of eighteen days took place in the same time period when my steps decreased dramatically. None of the other two managed to connect this trio of events.
In addition, Claude provided ranked recommendations for me rather than just listing them. First: rebuilding the movement, as this is the most glaring deficiency that separates the first half of my year from the second. Second: getting the anovulation issue addressed by a doctor as this is a medical issue and not one for tracking alone to solve. Third: HRV maintenance, as this is already positive.
It didn’t feel like I was reading a standard wellness document, but rather something that was written after consideration of my actual data.
How Each One Actually Interpreted My Data
This is definitely the part that made me wanna do this in the first place. All three had the same numbers before their eyes: same HRVs, same step counts, same six cycles with the same ovulation flags. However, what was important for each one individually and what they advised me to do with those numbers was different enough so that reading through all three one after another felt like hearing three separate opinions of three different doctors who analyzed the same chart.
Hormones and Ovulation: Same Flag, Different Weight
All three recognized my PMOS/PCOS history and the pattern of anovulation, three of six cycles with no evidence of ovulation and one cycle lasting 41 days (did anyone experience this March theory? I know many who did). But the importance each of them attached to this aspect of my situation was very different.
Gemini took it to be the crux of the problem and constructed almost all of its recommendations based upon it, insulin resistance being the cause, eating order being structured around protein and fiber first followed by carbohydrates and myo-inositol being a very specific request to my doctor. Quite narrowly and supplement-centric, as if it had considered managing PMOS/PCOS to be the only thing worth doing in the assignment.
ChatGPT recognized it as just one point among many instead of making it the main theme of the report. It placed the statement “ovulation confirmed in your two most recent cycles” under the list of positive changes and considered my anovulatory periods as part of history. In other words, Gemini implied this to be a persistent problem and ChatGPT implied that this used to be a problem and it was getting resolved.
Claude was the right combination of both. While it doesn’t state that there is any fix for the anovulation, neither does it present it only as the issue related to the supplementation. Instead, it takes into account the exact time frame of those anovulatory and atypical luteal phase cycles and compares it with the amount of steps I took within the same period of time, concluding that this is not an issue with stable hormones, but one which depends on my physical activity. Both Gemini and ChatGPT didn’t attempt to explain the reasons behind the occurrence of the issue.
Stress and HRV: The Same Number, Read Two Ways
My HRV increased substantially over the course of the last seven months and every tool was aware of it. However, the way it interpreted it varied depending on the time frame it selected for consideration.
Gemini considered my entire dataset average of 60.4 ms and designed its Stress section based on the correlation between my HRV and resting heart rate, negative 0.67 precisely. The interpretation given was general and related to my nervous system and the recommendations provided were general too. It recommended some regulation techniques, including breathing exercises and steady-state cardio training rather than sporadic intensive workouts.
ChatGPT concentrated on the latest trends and not the averages. It highlighted the fact that HRV had improved from the low forties and fifties throughout the year to the high seventies, 73-76 currently. This increase was singled out as the main achievement in the report without any further explanations for the reason behind it.
The last one, Claude, provided me with the most detailed information about my HRV. It reported 47 ms in December that rose to 78 ms in June. Moreover, this was the only tool that connected my achievements to the actions I should take to protect it going forward.
Movement: Who Connected the Dots and Who Didn’t
This is where the gap between the three was the widest. All three recognized that my daily step count was too low and recommended bringing it closer to 8,000 to 10,000 steps, providing practically the same advice.
The difference was in whether they found any connections between low activity and other data that I have. Gemini and ChatGPT considered physical activity as an independent factor, which was unrelated to any of the hormone and stress-related factors. In particular, the report of ChatGPT explicitly mentioned that the largest deficiency was not in my sleep, but in activity and nothing more.
Only Claude provided a connection between the physical activity and another issue. It noticed that my step count decreased from roughly 5,600 in the beginning of the year to 2,100 in May. In addition, it pointed out that the decrease coincided almost perfectly with the period when my cycle was anovulatory and the length of the luteal phase increased. Moreover, it recommended performing an experiment: increase my step count during the next two cycles and observe whether my ovulation and luteal phase will return to normal. Something I am genuinely making an effort to try out.
That said, my Ultrahuaman Ring did get a firmware update after which my step count has decreased drastically despite nearly the same movement throughout the day. Which, of course, none of the LLM models knew about.
Sleep: The One Section Everyone Agreed On
Sleep was the boring section in the best possible way. The three tools considered my numbers, an average score in the low 80s, efficiency of 91-92% and agreed on the same thing: this is good enough, do not fiddle with anything, do not strive for additional hours if not required.
Gemini referred to sleep as my “greatest health asset.” ChatGPT rated it with a perfect nine out of ten and advised me not to strive for nine or ten hours of sleep. Claude said pretty much the same, preserve the consistency, do not try to over-optimize something that already works. No real difference in opinion here, which in fact made me trust the three tools a little bit more as if they had come to completely different conclusions about the only one parameter that is definitely clear from the data.
Actionable Insights, Side by Side
Stacking the recommendations next to each other made the differences even clearer.
Gemini’s actions were clinical and supplements heavy. Myo-inositol, box breathing on certain intervals, meals timed for insulin. Good advice on paper but sounding like a PMOS/PCOS wellness course in general without any reference to the seven months’ worth of data I provided.
ChatGPT’s actions seemed to be the most well-structured and easily readable, especially the scorecard which rated me on a scale from one to ten in both sleep and movement aspects of the lifestyle. Nevertheless, the vast majority of individual actions sounded very generic and were aimed at such goals as walking more, eating more protein and strength training three times a week, with very little connection to my personal data.
Claude’s actions were sorted based on how important the information showed they should be, regardless of category. Rebuild your movement first since that’s the difference-maker in terms of changes from the first six to the second six months. Visit a doctor regarding the anovulation pattern second, since that problem requires the intervention of the professional who cannot be replaced by any tracking app. Protect the HRV gain third, because it’s already working. Every action pointed back to something specific in my numbers instead of reading like a template with my stats dropped in.
Why Not Go For Native AI Models?
Almost every wearable tech platform has a native AI built-in for the interpretation of the data it already collects. Oura has Oura Advisor, Ultrahuman has Jade, WHOOP has WHOOP Coach and many more. So why put my data through other LLM models?
You see, some of the apps limit the conversation you have with their AI models. Here’s looking at you, Jade! By inputting months of data into a separate model that doesn’t ask you to pay an additional fee, you can get as granular as you want with the number of questions you have about your data and health.
Another reason why I did this was the visual representation of my data. I already see the trend line graphs on my Oura app every morning and every week, small charts, one metric at a time, sitting on a screen the size of my palm. What I wanted was a different visual of the same numbers, something I could actually look at all at once instead of scrolling back six months on my phone just to see how one line moved from where it started to where it is now. I wanted to sit with the whole picture for more than five seconds, not swipe through it in pieces.
Setting This Up Yourself
If you are willing to experiment with your own wearable data, here are a couple of pieces of advice:
- Export Everything, Not Just the Summary- If you can export your raw daily CSV files from your device, use them, rather than whatever summary your app generates per week. The more months you have, the better any pattern it identifies will actually be a pattern and no mere coincidence.
- Specify the Request for a Dashboard- Instead of requesting insight or summary, mention the word ‘dashboard’, especially if the software allows building such and specify that you need it to be interactive.
- Provide Your Health Background- I informed the software about my PMOS/PCOS diagnosis and vitamin deficiencies and this was reflected in all further analysis. Otherwise, you will get only generic recommendations based on your wearable data.
- Check for Sense in Medical Recommendations- The good thing about all three tools is that all of them noticed my anovulatory cycles and advised me to visit a doctor. Of course, none of them diagnose and it would be very unwise to treat it like a tool that would. This could be only your starting point for discussion with your gynecologist.
Was It Worth Running the Same Data Through Three Tools
Yes, really. I would never know how each of them weighs the same figures in a different way if I had only tested one.
Moreover, I really appreciated the correlation between my movement and cycle timing, which is something I didn’t notice on a phone screen. Is my step count significantly lower since the firmware update? Yes. Could that have affected the LLM model as a sign of imbalance in my cycle and hormones? Most likely. But would I run this experiment again? 100%.
Since mine was a regular text prompt, if someone wants to make an easy-to-use dashboard for Claude wherein I only put in my data and it makes it a much better dashboard than what I created, please hit me up! And if you have done something like this with your wearable data, I would love to hear your experience. Let me know in the comments below!
