Abstract
Lithium has been used for decades as a mood stabilizer in bipolar disorder, but a growing body of epidemiological evidence suggests it may also confer neuroprotection — with some studies linking lithium exposure to reduced risk of dementia. The therapeutic implications are significant: lithium is inexpensive, widely available, and well-characterized pharmacologically. If a neuroprotective dose range could be established, it could be repurposed for cognitive protection in at-risk populations.
However, the existing evidence is far from conclusive. Most studies use binary exposure classifications (lithium yes/no), ignore dose–response dynamics, and fail to account for a critical confounder: reverse causation. Patients in the prodromal phase of cognitive decline may have their medications adjusted — including lithium dose changes — creating spurious associations between lithium levels and subsequent dementia diagnosis. Without accounting for this, any observed lithium–dementia relationship may be an artifact of clinical practice rather than a biological effect.
This study, conducted at Harvard Medical School and Brigham and Women’s Hospital, was designed to address these limitations head-on. Using the Mass General Brigham (MGB) Biobank’s longitudinal electronic health records, we built a complete analytical pipeline — from cohort construction (~15,000–20,000 patients) through data linkage, preprocessing, and survival modeling — with rigorous attention to the methodological pitfalls that have plagued earlier work. We modeled lithium as a time-varying continuous exposure rather than a binary variable, applied landmark analysis to mitigate reverse causation bias, and included appropriate confounders (age, sex, bipolar disorder status) from the start of the modeling process rather than deferring them to later stages. The pipeline also addressed multiple data quality challenges including EMPI missingness, post-diagnosis flag leakage, same-day test filtering, and appropriate gender column selection.
Preliminary findings reveal a U-shaped lithium–hazard relationship — suggesting that both very low and very high lithium levels may be associated with elevated risk, with an intermediate range conferring the greatest protection. The detailed results and their clinical implications are being prepared for publication.