Discussion

The goal of our analysis was to identify if there was a relationship between menstrual cycle and glucose levels of diabetic individuals. Our results from the Lomb-Scargle algorithm successfully demonstrated  a notable cyclic pattern in the glucose levels of young diabetic females and no cyclic pattern in diabetic males and older diabetic females (Results, Figures 3-4). Specifically, these patterns were found in key daily features of diabetes such as the 95th percentile value, minutes out of range, and percent time out of range. These metrics had  the smallest Chi-Square test p-values for both females vs. males and cyclic vs. acyclic females (Results, Table 1), indicating that there are significantly more women in their 30s with a high-power (value > 7.5) signal between 22 and 43 days. 

The 22-43 high-power signal cycle observed in the aforementioned features aligns with the typical 28-day menstrual cycle, suggesting that hormonal fluctuations throughout the cycle have a notable influence on metabolism and glucose regulation. Our findings align with previous work, as one study done by Herranz et al. (2016) found that 65.4% of their participants experienced cyclic changes with an increase in blood glucose levels from the early follicular phase to the late luteal phase. We hypothesize that the cyclicity in glucose levels results from hormonal fluctuations throughout the menstrual cycle which could impact metabolic processes as well as behavior. For example, during menstruation, the decline in estrogen levels can increase hunger and cravings, while also contributing to metabolic changes like insulin resistance. In fact, the three metrics with significant p-values across all comparisons were all related to high glucose values, which may be arising from the menstruation phase of the cycle. These physiological shifts throughout the menstrual cycle could be contributing to the cyclic fluctuations in glucose levels we see in our results. 

Furthermore, some of the most notable results were present in the T1D vs. T2D analysis. All six glucose metrics had significant p-values for the Chi-Square test, with the 95th percentile value on the order of 10-94, which is extremely significant. Even from visual inspection, we can see that although there was a small peak in the cyclic range of the T1D KDE, it was much smaller and less spread out than that of the T2D individuals. We expect this because T2D is much more influenced by behavior and other external features as compared to T1D, which is more physiological and often controlled earlier on (Sapra & Bhandari, 2023). This results in the greater variation and higher overall power of the T2D data, since those glucose signals are more prone to fluctuation. Overall, our results open up avenues for additional research based on DM type, looking at intersections of sex, age and, and treatment type. 

Additionally, we had some setbacks that put us slightly behind schedule for our analysis. First of all, it took a few weeks for Dexcom to transfer the data to our servers, and when we received it, it required a lot of preprocessing to get it in the correct format. Once we had the first version of our project complete, the servers hosting our data and code were corrupted, causing us to lose our work. We had to preprocess the data again and write our scripts from scratch. Therefore, at the time of writing, we haven’t been able to finish the machine learning portion of our project, and there are some additional analyses for T1D vs. T2D that we won’t have time for, however we will continue analyses in the future. 

Overall, our results have meaningful implications; as shown in the figures above, cyclicity has influence on mechanisms such as metabolism and glucose-regulation, whether in diabetics or healthy patients. These findings should guide current algorithm designs for CGM or other glucose-tracking/prediction devices to implement menstrual cyclicity as a metric; that way, data from cyclic females can be more accurately interpreted. Taking cyclicity into account can improve diagnosis and treatment for diabetic females and could potentially combat unsafe or ineffective treatments. Additionally, this study focused on menstrual cycle impact on label-free glucose data, but there are other modalities such as hormone levels and behavior that can also improve our understanding of the interplay between cyclicity and DM. Future directions we would like to take our study include identifying clusters of glucose features related to cyclicity, continuing T1D vs. T2D analysis, and ensuring reproducibility with diverse populations across the globe. 

Leader: Safa Saeed