TL;DR
This study uses a large US claims database to investigate how the presence of various health conditions (comorbidities) relates to mild cognitive impairment (MCI), finding that this relationship is strongly dependent on age. The research demonstrates that a machine learning model using these comorbidities can predict MCI with reasonable accuracy, particularly in younger individuals, suggesting a potential tool for early risk identification.
What Problem Does This Paper Address?
Mild cognitive impairment (MCI) is a crucial transitional stage that can precede Alzheimer's disease and related dementia disorders (ADRD). Early identification of individuals with MCI who are at higher risk of progressing to ADRD is vital for developing personalized prevention strategies and facilitating timely treatment. However, current methods for identifying MCI in primary care settings are limited, facing barriers such as a lack of widely used assessment tools, insufficient training, and time constraints during patient visits. This paper aims to address this gap by exploring whether existing health conditions (comorbidities) commonly associated with ADRD can serve as predictive markers for MCI, and how this predictive power changes across different age groups.
Methods
The researchers employed a retrospective matched case-control design using the Merative MarketScan Commercial and Medicare Databases. They identified two cohorts: individuals with an MCI diagnosis but no ADRD diagnosis at entry, and a control group without MCI or ADRD. These cohorts were matched based on age, sex, geographic region, and the year of data entry using a propensity score method.
A key aspect of their approach was the use of machine learning for MCI risk prediction. Specifically, they utilized a Bayesian logistic lasso regression (BLLR) model, chosen for its simplicity and interpretability, which performed comparably to other tested algorithms like stochastic gradient boosting and XGBoost. The BLLR model incorporated age, sex, and 25 identified ADRD comorbidities as predictors. The model's performance was evaluated using the area under the receiver operating characteristic (ROC-AUC) curve.
Datasets
The study utilized the Merative MarketScan Commercial and Medicare Databases. These databases contain health services data for employees, dependents, and retirees in the US with primary or Medicare coverage. The observation period spanned from January 1, 2014, to December 31, 2019. This dataset is representative of a significant portion of the US population with private insurance or Medicare.
Key Insights and Ideas
- Comorbidities as MCI Predictors: The study strongly suggests that conditions often linked to ADRD are also significant comorbidities for MCI itself, not just its progression. This broadens our understanding of the interconnectedness of various health issues and cognitive decline.
- Age is a Crucial Moderator: The most striking finding is how the predictive power of comorbidities for MCI risk is heavily influenced by age. The association between comorbidities and MCI is weaker in older age groups, meaning a condition that strongly predicts MCI in a younger person might have a much lower predictive value in an older person.
- Younger Age Groups Show Stronger Prediction: The machine learning model performed best in predicting MCI for the youngest age group (50-64 years), indicating that comorbidities might be more indicative of early-stage MCI risk before other age-related factors become dominant.
- Specific Comorbidities Emerge as Key: Certain comorbidities, like depression, stroke/TIA, obstructive sleep apnea, and hearing loss, consistently appeared as significant predictors across multiple age groups, highlighting their potential as core indicators.
- Potential for a Screening Tool: The findings support the development of a practical screening tool for primary care physicians (PCPs) that leverages readily available EHR data (like comorbidities) to flag individuals at higher risk for MCI, even if they aren't actively reporting symptoms.
Key Results
- A total of 5185 MCI individuals and 15,555 matched non-MCI individuals were analyzed.
- The overall model (BLLR) achieved an ROC-AUC of 0.72 for predicting MCI.
- Age-Specific Performance:
- Age 50-64 years: ROC-AUC of 0.75 (Sensitivity: 61.9%, Specificity: 77.4%)
- Age 65-79 years: ROC-AUC of 0.70 (Sensitivity: 59.3%, Specificity: 71.7%)
- Age ≥ 80 years: ROC-AUC of 0.66 (Sensitivity: 55.3%, Specificity: 66.8%)
- The odds ratios (ORs) for comorbidities predicting MCI decreased significantly with increasing age. For example, depression had an OR of 4.4 in the 50-64 group, dropping to 2.9 in the 80+ group. Stroke/TIA showed a similar trend, with ORs of 6.4 and 2.1, respectively.
- Key Predictors by Age Group:
- 50-64 years: 12 comorbidities were significant predictors, including depression, stroke/TIA, obstructive sleep apnea, hearing loss, bipolar, psychosis, hypothyroidism, insomnia, weight loss, chronic pulmonary disease, alcohol abuse, and drug abuse.
- 65-79 years: 5 comorbidities were significant: stroke/TIA, depression, obstructive sleep apnea, hearing loss, and weight loss.
- 80+ years: 4 comorbidities were significant: depression, hearing loss, stroke/TIA, and weight loss.
Limitations and Future Work
The authors acknowledge several limitations:
- Claims Data Limitations: Claims data lack detailed clinical variables (e.g., body mass index, smoking status, severity of conditions) and cannot fully assess the nuances of comorbid conditions.
- Generalizability: The study population is drawn from commercially and Medicare supplementally insured individuals, potentially limiting generalizability to uninsured or underinsured populations.
- Underdiagnosis of MCI: The analysis may underestimate the true burden of MCI due to underdiagnosis, particularly in older age groups where the prevalence of undiagnosed MCI might be higher.
- Observational Design: The retrospective, case-control design precludes establishing causality.
- Temporal Bias: The case-control design might introduce bias, as individuals who see a doctor more often might be more likely to be diagnosed with MCI.
Future work could address these limitations by:
- Utilizing Electronic Health Record (EHR) data, which offers a broader range of predictors and can potentially differentiate MCI due to AD from general MCI.
- Conducting longitudinal studies to better understand causality and the temporal relationship between comorbidities and MCI.
- Comparing individuals with diagnosed MCI to those with clinically verified normal cognition.
- Integrating blood-based biomarkers into predictive models as they become more mature.
- Developing and validating screening tools for PCPs to identify high-risk individuals for further assessment.
Summary updated