Background and Motivation

Over the last years, many studies and headlines have suggested that COVID-19 might increase the risk of neurodegenerative diseases (NDDs) such as Alzheimer’s Disease (AD) and Parkinson’s Disease (PD). Several meta-analyses (e.g., [1], [2]) report a higher risk of receiving an NDD diagnosis after COVID-19, but they are almost all based on observational data (like Electronic Health Records (EHRs)) and standard statistical adjustments.

In our recent paper, we revisit this question using causal inference and modern machine learning—and arrive at a more nuanced picture:

• We do not find strong evidence that COVID-19, on average, causally increases the risk of being diagnosed with a neurodegenerative disease within about five years.
• We do find vulnerable subgroups where the risk appears to be substantially higher, consistent with COVID-19 unmasking or accelerating existing disease processes rather than causing NDDs de novo.
   

Why associations from EHR data can be misleading

Most previous studies compare NDD incidence between “COVID-positive” and “COVID-negative” patients in EHRs. Even with propensity score matching of negative to positive patients, this setup is vulnerable to several biases:

• Baseline confounding: People who get infected or tested differ in age, comorbidities, socioeconomic status, and other factors.
• Selection bias and changing testing policies: Early in the pandemic, testing mainly targeted high-risk groups; later, testing broadened.
• Censoring bias: People infected earlier simply have more follow-up time, which can inflate their observed risk.
• Competing risk of death: Many dementia risk factors also increase mortality. Death can “compete” with an NDD diagnosis and distort naïve estimates.

If these issues are not modelled explicitly, an association between COVID-19 and NDD can easily be over- or underestimated.
 

Our approach: Doubly robust causal time-to-event analysis

We frame the question in terms of causal effects: what is the cumulative incidence of NDD diagnoses within a given time horizon if everyone had COVID-19 versus if no one did? These are counterfactual risks that we estimate under standard causal assumptions.

Methodologically, we use Targeted Maximum Likelihood Estimation (TMLE) for survival and competing risks [3], implemented in our open-source Python package PyTMLE [4].

In short, we:

• Fit flexible ML models (gradient-boosting models) for:
  • The hazard of NDD (AD, PD, unspecified dementia, or a composite endpoint summarizing all three),
  • The hazard of death,
  • The censoring mechanism,
  • The propensity of being COVID-positive.
• Use TMLE to target these models toward the causal estimand: the difference or ratio of cumulative NDD incidences with vs. without COVID-19.
• Obtain doubly robust estimates: if either the exposure/censoring models or the outcome models are correct, the causal effect estimates remain consistent.

Our data stems from three large-scale collections of EHRs from Denmark (Danish National Patient Registry and others, N_total~8,000,000), the US (provided by TriNetX [5], N_total~3,000,000), and the UK (UK Biobank [6], N_total~500,000). In the Danish and US data, effects can be targeted at time horizons of up to 5 years. Our outcomes of interests are the first reported diagnoses of AD, PD, dementia, or a composite endpoint summarizing all three.

Main findings

1. No average causal effect in the considered populations
Across cohorts, waves, endpoints, and both control designs, average treatment effects (ATEs) on NDD risk are small and statistically non-significant once we properly adjust for confounding, censoring, and competing risk of death.

When we use historical controls, estimated effects are sometimes larger and occasionally significant. However, sensitivity analyses indicate that these “signals” could plausibly be explained by unobserved, time-dependent confounders (e.g., effects of lockdowns or changes in health care). In contrast, analyses with contemporary tested-negative controls are more robust and consistently show no significant causal increase in NDD diagnoses.

Our risk ratio estimates are generally below those reported in recent meta-analyses (e.g., [1], [2]), even though we include up to five years of follow-up.

2. Vulnerable subgroups with elevated risk

Average effects can hide important heterogeneity. In addition to the TMLE-based estimates, we derive individual-level counterfactual risks and inspect individualized treatment effects (ITEs) – how much each person’s NDD risk changes under “COVID” vs. “no COVID”.

• For the majority (~90 %) of patients, the estimated risk difference is close to zero.
• However, in the top ~5 % with the largest ITEs, we see substantial risk increases (e.g., >5–15 % absolute increase in parts of the TriNetX cohort).

These high-risk patients tend to be:

• Older,
• With more prior diagnoses and more recent health-care encounters,
• Enriched for known NDD risk factors: type 2 diabetes, cardiovascular disease, depression, hearing impairment, obesity, etc.

This pattern supports a model where COVID-19 accelerates or unmasks pre-existing undetected neurodegenerative pathology in vulnerable individuals, rather than causing a large population-wide shift in NDD incidence.

From a public-health perspective, our results are cautiously reassuring: we do not see a strong, population-wide causal increase in neurodegenerative diagnoses after COVID-19 in the first years after infection. At the same time, the consistent signals in high-risk subgroups highlight an important direction for future work: predicting which individuals are most vulnerable, and why.

Project COMMUTE

References

[1] D. Shan, C. Wang, T. Crawford und C. Holland, „Association between COVID-19 infection and new-onset dementia in older adults: a systematic review and meta-analysis,“ BMC Geriatrics, 2024. DOI: 10.1186/s12877-024-05538-50

[2] Q. Zhang, R. X. Y. Botta, J. C.-C. Wei und T.-H. Tung, „Risk of new-onset dementia following COVID-19 infection: a systematic review and meta-analysis,“ Age and Ageing, Bd. 54, Nr. 3, p. afaf046, 2025. DOI: 10.1093/ageing/afaf0463

[3] H. Rytgaard und M. van der Laan, „Targeted maximum likelihood estimation for causal inference in survival and competing risks analysis>,“ Lifetime Data Analysis, Bd. 30, Nr. 1, pp. 4-33, 2024. DOI: 10.1007/s10985-022-09576-20

[4] J. Guski, M. Aborageh und H. Fröhlich, „PyTMLE: A Flexible Python Library for Targeted Estimation of Survival and Competing Risks using Causal Machine Learningq,“ medRxiv, 2025. DOI: 10.1101/2025.07.02.25330730

[5] M. B. Palchuk, J. W. London, D. Perez-Ray, Z. J. Drebert, J. P. Winer-Jones, C. N. Thompson, J. Esposito und B. Claerhut, „A global federated real-world dat and analytics platform for research,“ JAMIA Open, Bd. 6, Nr. 2, 2023. DOI: 10.1093/jamiaopen/ooad0353

[6] C. Sudlow, J. Gallacher, N. Allen, V. Beral, P. Burton, J. Danesh, P. Downey, P. Elliott, J. Green, M. Landray, B. Liu, P. Matthews, G. Onk, J. Pell, A. Silman, A. Young, T. Sprosen, T. Peakman und R. Collins, „UK Biobank: An Open Access Resource for Identifying the Causes of a Wide Range of Complex Diseases of Middle and Old Age,“ PLoS Medicine, Bd. 12, Nr. 3, pp. 1-10, 2015. DOI: 10.1371/journal.pmed.1001779