The results of this retrospective study of the MarketScan Commercial and Medicare Databases suggest that ADRD comorbidities are also comorbidities for MCI. The BLLR algorithm was selected due to its simplicity and interpretability and because it yielded similar prediction performance as the other machine-learning algorithms.

The 25 ADRD comorbidities identified by our literature search were also significant risk factors for MCI in this population. Individuals in the MCI cohort had a higher frequency of comorbidities compared with the non-MCI cohort. The differences between cohorts for depression and stroke/TIA were the largest. Depression, stroke/TIA, obstructive sleep apnea, insomnia, hearing loss, ischemic heart disease, and hyperlipidemia appeared on both the list of the highest frequency comorbidities and the list of comorbidities with the highest ORs, suggesting these seven comorbidities may be predictive of MCI. Considering that MCI is a prodromal stage for ADRD, the connection between ADRD comorbidities and MCI comorbidities is not unexpected; however, identifying the comorbidities with the strongest connections and quantifying the relationships can help inform the development of a screening tool to identify high-risk individuals.

Secondly, the ORs (MCI vs. non-MCI) decreased with increasing age group for all comorbidities. The differences were statistically significantly higher for the age 50 to 64 group compared with both older age groups. As a result, the potential for ADRD comorbidities to predict MCI risk also declined with increasing age group. The BLLR results in this study demonstrated better model predictivity in the younger age group. Depression, stroke/TIA, hearing loss, and weight loss were significant predictors across all age groups; however, obstructive sleep apnea was significant only for the two youngest age groups. Furthermore, hypothyroidism, insomnia, bipolar, chronic pulmonary disease, psychosis, alcohol abuse, and drug abuse were only significant for the youngest age group.

Changes in the likelihood of observing a particular comorbidity in different age groups may be related to the epidemiology of the specific condition. For cardiovascular and metabolic diseases, frequencies increased by age group in both the MCI and non-MCI cohorts; thus, ORs decreased. This aligns with the literature because people are more likely to develop cardiovascular and metabolic diseases with increasing age [35,36,37]. The association between hypertension in midlife, which aligns with the youngest age group for this study, and cognitive decline has been well established [38, 39]. The relationship between hypertension developing in late life and dementia is less clear. A 24-year prospective study found that in addition to hypertension in midlife and late life, a history of hypertension followed by late-life hypotension was also associated with an increased risk of dementia [40]. Our study evaluates hypertension over a shorter time period but confirms the importance of considering the temporal effects of comorbidities.

The frequency of psychiatric disorders, including depression and bipolar, decreased in the oldest age group in both cohorts. This aligns with the literature because the prevalence of psychiatric disorders decreases with age, in part because of the reduction in life expectancy of people with depression [41].

The observed differences in the oldest and youngest age groups in this study may have been impacted by survivorship bias. To reach the oldest age group, individuals likely maintained good health during the prior years. The number of years with comorbidity is likely to impact the onset of dementia and longevity [42].

Another possible contributor to lower ORs in the older age groups in this study may be the effect of undiagnosed MCI in the non-MCI cohort. The proportion of undiagnosed MCI individuals is likely similar to MCI prevalence at the population level, which increases with age. For people aged 50 to 64 years, the MCI prevalence is about 6.7% [4]. If our cohorts are similar to the general population, the non-MCI cohort may include a similar percentage of undiagnosed MCI individuals, which should have little impact on the frequencies for comorbidities in the non-MCI cohort and the ORs. However, for the 80+ years group, the MCI prevalence in the general population is about 25.2% [4]; the impact of undiagnosed MCI on ORs in this age group may not be ignored.

Our findings suggest that chronic pulmonary disease is a predictor of MCI; however, reports of the association between chronic obstructive pulmonary disease and ADRD outcomes are conflicting and limited [35, 36]. Thus, additional research on the association between chronic pulmonary disease and MCI and ADRD is recommended, as well as further evaluation to quantify the impact of smoking on MCI.

Chronic periodontitis and head injury were excluded from this analysis because fewer than 0.3% of individuals received a diagnosis. For chronic periodontitis, this may be because dentists were most likely to diagnose and treat that condition. MarketScan data do not capture dental records. A diagnosis of chronic periodontitis would only be captured in the dataset if it were noted in a medical office setting. For head injury, a longer baseline period may be required to quantify the relationship.

According to the package inserts for aducanumab and lecanemab-irmb, treatment should be initiated in patients with MCI or at the mild dementia stage of disease [14, 15]. With the availability of new treatments and screening tools for mild and early-stage ADRD, clinical guidelines will need more frequent updates that describe best practices for people with ADRD [1, 43]. Diagnosis of MCI due to AD is important to patients and their families, providing opportunities for treatment and future preparations. To have the greatest impact, predictive models should focus on identifying individuals at elevated risk for MCI.

A predictive model for MCI risk based on EHRs could include demographic characteristics (e.g., age, sex, race/ethnicity), biometric data (e.g., blood pressure, body mass index), health-related behavior (e.g., smoking status), laboratory results (e.g., lipids, HbA1c), and the presence of ADRD comorbidities and other data that are accessible in the PCP setting. The model will not include biomarkers, and thus, it will not be a diagnostic tool. However, an easily implemented screening tool for PCPs can greatly improve their ability to identify individuals at elevated risk of MCI. Alerting the PCPs of the possibility of undetected risks could provide an entry point for triage when an individual is flagged for elevated risk. Depending on the maturity of blood-based biomarkers, PCPs could use those results as part of their initial work-up and to decide whether and with what urgency to initiate specialist referrals.

One of the linchpins in the pursuit of the early detection of MCI by PCPs in age-eligible patients is the Medicare Annual Wellness Visit (AWV). Beginning in 2011, the AWV includes the detection of cognitive impairment for Medicare Part B beneficiaries [44]; however, by 2018 uptake of the AWV was still only at 32% [45]. Medicare is primarily available to people aged 65 years or over and this age group is considered the most at risk for MCI and dementia [46]. Being able to identify individuals at risk for MCI before they reach 65 years would enable physicians to treat and track them even earlier, thus potentially limiting the clinical and economic burden of the progression to ADRD.

The results of this study must be considered in the context of several limitations. Firstly, there are those inherent to all claims data: claims data do not allow for proper assessment of potentially relevant clinical variables such as body mass index, smoking status, and the severity, rather than mere presence, of comorbid conditions. Additionally, the generalizability to populations other than the commercially and Medicare supplementally insured, also referred to as Medigap, is unknown. Data from MarketScan is sourced from employers; findings may not be generalizable to the uninsured or underinsured populations. Claims data are collected for reimbursement and not research purposes. This limitation can be addressed by future studies using EHR data, which contains a broader range of predictors and covers an “all-comers” population. An additional benefit to conducting a similar study using EHR data would be the ability to separate individuals with MCI due to AD from the general MCI population, something that was not possible in this study. As mentioned above, because of the underdiagnosis of MCI, this analysis may underestimate the true burden of MCI. The pattern of odds ratio decreasing with increased age may be partially attributed to the expectation that the highest percentage of undiagnosed MCI individuals in the non-MCI cohort is likely to be in the oldest age group. While longitudinal, our observational study design precludes the assessment of causality. Increased diagnoses related to complications as patients near an AD dementia diagnosis have been documented in the literature, which may reflect increased use of health care services as cognitive impairment worsens [47]. Studies that advance our understanding of the diagnostic process, as well as the natural history of the AD continuum, may further elucidate the relationship. In order to address these limitations, studies that compare individuals with diagnosed MCI to individuals with clinically verified normal cognition are needed. In addition, the potential temporal bias could be introduced by using a case-control study design, e.g., patients who saw a doctor more often are more likely to be diagnosed with MCI.