In recent years, scientific research has significantly expanded knowledge in the field of diabetology. Substantial changes in therapeutic practices have been coupled with significant improvements in the monitoring of diabetic individuals.

Compared to traditional measurements, which primarily focus on individual blood glucose values rather than the overall glycemic trends over a defined time frame, alternative glycemic profiles have gained prominence. These include “Time In Range” (TIR), which is associated with “Time Above Range” (TAR) and “Time Below Range” (TBR), representing the time period during which diabetic individuals maintain blood glucose values within predetermined normal ranges [24]. Other methods include the “Ambulatory Glucose Profile” (AGP), which involves continuous glucose monitoring for 24 h over a specified period (typically two weeks), glycemic variability, which tracks the frequency and amplitude of glycemic fluctuations, and the coefficient of glycemic variation (CV), which correlates the standard deviation (SD) of measured blood glucose values with the mean blood glucose level [25]. Among these methods, glycemic variability and CV undoubtedly offer a more effective assessment of the “glycemic health” of individuals with T2DM.

Back in 1970, Service et al. associated the concept of glycemic variability with glycemic excursions typically induced by meals, introducing the concept of the Mean Amplitude of Glycemic Excursions (MAGE), using glycemic excursions exceeding a single standard deviation (SD) above the mean glucose level [26]. However, it was only in 2003, thanks to a paper by Kovatchev et al., that the link between increased hypoglycemic events and high glycemic variability was established in both type 1 and T2DM [27].

In 2005, Hirsch, who developed a formula based on SD and mean glucose levels to assess if a series of glucose values followed good glycemic control, and Brownlee hypothesized that inconsistent glycemic control with high glycemic variability underlies the development of microvascular complications in diabetes, more so than chronic hyperglycemia [28]. In 2006, Monnier et al. suggested that these complications were more associated with high levels of oxygen free radicals induced not so much by chronic hyperglycemia but by large glycemic fluctuations [29]. In contrast, a few months later, Garg and colleagues demonstrated the significant benefits of continuous glucose monitoring (CGM) techniques, already within 10 days of use [30].

As mentioned above, less literature attention has been devoted to the relationship between the levels of average blood glucose sustained in the period preceding a hospital admission (in our case, for medical rather than surgical reasons) and in-hospital blood glucose levels. The approach to patients with T2DM during hospitalization is almost standardized, aiming to achieve a glycemic target that typically falls within the same ranges.

In most cases, oral antidiabetic therapies are suspended at hospital admission, and patients are transitioned to subcutaneous insulin administration using a basal-bolus regimen. There are exceptions for patients admitted to intensive care settings (where higher average glucose levels may be allowed due to acuity that benefits from more readily available resources), critically ill patients, and those considered more clinically stable: the latter group may be recommended to continue or resume home-based therapies also during hospitalization.

The primary objective of this work was to investigate the influence of pre-hospital admission blood glucose levels on the outcomes observed during the hospitalization period: intuitively, accessing these data is not straightforward, particularly when dealing with individuals who do not use continuous home monitoring techniques such as CGM or Flash Glucose Monitoring (FGM). One way to overcome this challenge is to use HbA1c testing at the time of hospital admission, which is already strongly recommended by current guidelines. To obtain the estimated average pre-hospitalization blood glucose, a simple conversion table can be used, based on the linear correlation between HbA1c levels and the average blood glucose [31].

An initial approach to characterize our patient population relied on the use of ML techniques to generate a predictive model of 30-day mortality from data collected during hospitalization. As shown in the previous sections, AdaBoost-FAS, the method we designed for our cohort, displayed impressive predictive abilities: for example, it was able to predict the 30-day mortality class of diabetic individuals with a balanced accuracy on the independent test set of 85.6%. The subsequent reliance analysis performed showed that, overall, the model is particularly sensitive to changes in two specific variables, namely the exclusive home treatment with SGLT-2 inhibitors (reliance = 4.2) and a positive history of non-metastatic hematological or solid neoplasms (reliance = 3.6). However, overall, the model predictions are also fairly sensitive to changes in 24 other variables (reliance > 2.7). This suggests that it is not possible to reliably determine the 30-day mortality risk based on measuring only a handful of variables. Therefore, providing an intuitive explanation of the decision-making process of the model to clinicians appears challenging and the potential lack of transparency may limit the trust clinicians place in the model, subsequently reducing the likelihood of its adoption.

To further discriminate the factors most associated with 30-day mortality in our patients, we decided to use classical statistics, including population descriptive analyses, logistic regressions, and survival curves, stratifying the population by outcome (death yes/no), levels of glycemic variability (SD < or ≥ 42.55 mg/dl), glycemic CV (< or ≥ 25.8%), and estimated pre-hospitalization glycemic CV (< or ≥ 28.8%). This last variable is not found in the literature and has been introduced by our research group: it consists of a ratio between glycemic variability recorded for our patients during hospitalization (expressed in mg/dl) and the pre-hospitalization mean glucose calculated from the HbA1c levels of diabetic individuals, tested necessarily within the first 48 hours since hospital admission (also this last variable is expressed in mg/dl).

The rationale behind this calculation is to be found in the original purpose of the study, which is precisely to investigate the existence of a relationship between home blood sugar levels and those recorded during hospitalization. The fact that the numerator and denominator are expressed in the same unit of measurement also reduces the risk of over- or underestimating the value of individual variables.

From the analysis of our population, it emerged that there is a significant difference in the percentage of diagnoses of dementia or hemiplegia between patients who died within the first 30 days of hospitalization and those who survived. Patients with worse outcomes often had at least one of the diagnoses of dementia and hemiplegia upon admission to our department, and these factors have proven to be the most reliable in predicting 30-day mortality, even in logistic regression analyses. This is in line with current literature, and in 2022, a group of researchers from our university demonstrated on a dataset of over 3 million patients admitted to Italian Geriatrics and Internal Medicine Units that dementia is among the diagnoses most predisposing to in-hospital mortality [32]. The same occurs in the case of hemiplegia, an outcome in our patients often following a prior stroke, which, in turn, is a negative prognostic factor for diabetic and non-medically hospitalized subjects for all causes [33, 34].

The explanation behind this phenomenon is intuitively connected to the initial clinical conditions of patients with dementia or residual hemiplegia, who are considered intrinsically fragile and more predisposed to further acute events, such as additional cerebrovascular events, infections, or new cancer diagnoses. Therefore, it is the performance status of these subjects, combined with a limited organic response to acute events, that leads to an exitus in the short term of hospitalization.

Also, from population descriptive analyses, it emerged that both the subgroups of patients who died/survived within 30 days of hospitalization and those with intra-hospital blood glucose CVs greater or less than the threshold value of 25.8%, showed significantly different lengths of stay. In particular, both the patients who survived to the 30th day and those with higher glycemic CVs experienced longer hospital stays. This demonstrates that diabetic subjects encountered worse outcomes, especially in the early stages of hospitalization, while those with longer hospital stays more frequently experienced higher fluctuations in blood glucose, suggesting the difficulty on the part of the clinicians in achieving adequate glycemic control in patients with a diagnosis of T2DM.

Similarly, significant differences between groups (population stratified by SD and intra-hospital blood glucose CV) were found in terms of HbA1c levels. Once again, higher levels of glycated hemoglobin (and therefore average pre-hospital blood glucose levels) were found in subjects who experienced greater blood glucose fluctuations during hospitalization (both SD and CV), further demonstrating that poor at-home glycemic control can (not only theoretically) predispose to equally inadequate in-hospital glycemic control.

The data related to blood glucose levels, which vary significantly between patient subgroups with different SDs and CVs both during and before hospitalization, are straightforward and therefore do not require further comments. It is worth noting that the “pure” data regarding mean and median blood glucose levels during and before hospitalization did not reach statistical significance in distinguishing between patients who died and those who survived on the 30th day of hospitalization.

Regarding another noticeable difference in Table 1, namely the percentages of subjects with COPD in the subgroups stratified by SD and intra-hospital blood glucose CV, it can be hypothesized (although not proven, as no checks were conducted on medical treatments administered during hospitalization) that among the reasons for hospital admission for these patients, there may have been an illness needing treatment with steroid medications (such as exacerbations of chronic pulmonary diseases) and, therefore, a cause of significant increases in blood glucose levels during the hospital stay. This would be in line with current literature and would justify the inadequate glycemic controls observed in some subpopulations [35].

The subgroups of the population stratified by death/survival within 30 days also differed in terms of the medical history component related to metastatic tumoral disease, with nearly double percentages in the deceased population. This could be explained in the same way as described for the variables “hemiplegia” and “dementia,” although it cannot be denied that the small sample size (4 subjects per subgroup had the same variable) in this case may have overestimated the difference.

As for the differences in terms of at-home treatment, the data should be interpreted similarly to what was mentioned earlier regarding HbA1c; substantial differences between groups were observed for dietary therapy, which was more prevalent in the subgroup with lower SD and intra-hospital blood glucose CV, and insulin therapy, with the opposite trend. In this case as well, it can be hypothesized that patients treated with milder therapies, such as dietary therapy, and presumably accustomed to achieving lower blood glucose levels at home (unlike those treated with insulin), were somehow less prone to metabolic imbalances even during hospitalization.

Regarding the analysis shown in Table 2, there is not much to say: in our case, the estimated pre-hospital CV is the only variable to achieve a statistical significance in association with 30-day mortality and it demonstrates, at least in our patient cohort, a possible theoretical correlation between pre-hospital mean glycemia and hospitalization outcomes. Whether this will have a real-world counterpart remains to be fully demonstrated in other studies with larger sample sizes, since the introduction of the variable in the literature was made for the first time in this pilot study.

The logistic regression analyses have been partly discussed, at least regarding the predictive role of “hemiplegia” and “dementia” in 30-day mortality. In these analyses, age, sex (male sex was used as reference), and length of stay in days were always considered as potential variables determining better or worse outcomes. The first of the three regressions showed that, after hemiplegia and dementia, the variables sex and SD ≥ 42.55 mg/dl are capable of independently influencing 30-day mortality in our patients (OR 7.64 and 5.49, respectively). In both cases, the data is highly consistent with current literature: male sex is known to be the foremost unmodifiable risk factor for in-hospital mortality and predetermined time intervals after admission [36, 37]. Similarly, there are no particular doubts about the deleterious effects of high glycemic variability (expressed as SD) during hospitalization. A 2017 study by British researchers revealed a strong association between high glycemic variability and increased in-hospital mortality in a population of over 28,000 subjects [38], and a 2022 Korean study conducted on over 2,000 patients (diabetic and non-diabetic) admitted for acute heart failure showed increased one-year mortality in subjects with a higher number of glycemic fluctuations during hospitalization [39].

Continuing with our analyses, a second logistic regression model, which substituted the variable intra-hospital CV ≥ 25.8% for SD, showed results consistent with those recorded in the first regression. Specifically, after hemiplegia and dementia, both sex and length of stay were strongly predictive (OR 8.64 and 0.86, respectively), with intra-hospital glycemic CV ≥ 25.8% following closely with an OR of 7.53 (p = 0.028).

It is significant that there is only one comment concerning this last variable: as seen, the data about glycemic variability expressed as SD could have a role in determining hospitalization outcomes, and the glycemic CV, which also provides information regarding the mean glucose levels recorded within a time frame, cannot be any less important. This has been suggested by recent observations in a home setting and allows the CV to play a key role in the extra-hospital glycemic monitoring of diabetic subjects. In particular, a 2019 consensus, also adopted as a model by the Italian Society of Endocrinology (SIE) [40], established that a CV < 33% is able to minimize home hyper- and hypoglycemic events, as well as protect diabetic subjects from chronic complications of T2DM, especially microvascular ones, as highlighted in a study by Hirsch et al. back in 2015 [41].

A significant degree of predictability for 30-day mortality from hospital admission is also attributed to the variable our research group introduced in this study (estimated pre-hospitalization glycemic CV), which was introduced in the third logistic regression analysis alongside the previously described variables. In terms of predictive strength, this index appeared after hemiplegia, dementia, and sex, showing an OR of 4.84 (p = 0.042), and showed that the relationship between intra-hospital glycemic variability and average pre-hospitalization blood glucose levels (derived from HbA1c) is closely associated with 30-day all-cause mortality in patients with T2DM.

It should be noted that, in all three analyses, the length of stay was the only variable that presented a “protective” effect against 30-day mortality. However, this data has been previously discussed and does not deserve further comments.

To provide a more in-depth estimate of the effectiveness of the variables under consideration, three different Cox regression analyses were conducted on our population of diabetic patients, adjusting for age, sex, and comorbidities (using the CCI). These analyses showed that, at 30 days, all subgroups of patients with worse glycemic control (values of SD, intra-hospital glycemic CV, and estimated pre-hospitalization glycemic CV above the calculated cut-off) had lower survival rates; however, for the variable intra-hospital glycemic CV it can be observed that the survival curves of subjects with values below and above the 25.8% cut-off intersected in the first days of hospitalization, indicating a poor discriminatory capacity between populations with better or worse outcomes, at least for that specific period of stay.

As a whole, the study seems to confirm both known aspects in the literature, such as the association between high glycemic fluctuations and increased 30-day mortality in patients with a diagnosis of T2DM, and it reveals a completely original aspect that we believe might be considered during the initial in-hospital evaluation of these subjects. The association between average blood glucose levels, both during hospitalization and pre-hospitalization (the latter easily obtainable through HbA1c measurement), and the degree of intra-hospital glycemic variability seems to have a high capacity to discriminate patients at a higher risk of dying within 30 days from hospital admission for all causes. Pre-hospitalization average blood glucose, when compared with intra-hospital glycemic variability (placed in the denominator), shows an even stronger association than what was calculated during hospitalization, suggesting, if it is still needed, that measuring HbA1c is useful and necessary for the correct evaluation of diabetic patients.

The limitations of this study are numerous, primarily due to its retrospective nature, which does not allow for any interventional changes to be made to the patients during the course of the study. It should be emphasized, as previously mentioned, that no information regarding the reasons for hospital admission for the patients has been reported (only surgical causes were excluded), nor information about the treatments administered to them during hospitalization.

The small sample size (120 patients, calculated based on the percentage of diabetic patients hospitalized per month in our department) and the single-center nature of the study do not allow for any definitive conclusions to be drawn about the observations made. Furthermore, the absence of continuous glucose monitoring (or flash monitoring) in favour of point-of-care testing (POCT) certainly does not provide the same level of precision in glycemic sampling and is influenced by the methodology in use.

We are aware, therefore, that a significant expansion of the analytical sample is necessary to overcome, at least in part, the limitations described. Similarly, the definition of a standardized research protocol and the involvement of a larger number of participating centers could lead to significant discoveries in this area of in-hospital management of diabetic patients, which is still largely unexplored.