Project Summary
Methods to characterize patients who highly bene¿t on multiple clinical outcomes, from treatments in Alzheimer's
disease and related dementias (ADRD), are necessary to treat patients effectively. Treatments may bene¿t some
patients on targeted outcomes, but harm some patients on other, e.g., cognitive, outcomes. So, characterizing
patients who highly bene¿t on multiple outcomes is signi¿cant: ¿rst it allows these patients to choose a treatment
if it is predicted to give them high bene¿ts without the harms; second, accurate characterization methods do not
exist. Generally, a characterization method has two stages. One stage “constructs” outcome predictions based
on patients' covariates; and another stage “synthesizes” the predictions to estimate the goal - a large high bene¿t
group. As the “construction” stage uses many covariates, it needs methods to estimate predictions from a model,
i.e., from a large set of possible distributions (e.g., regression, neural networks). These predictions are then used
in the “synthesis” stage for the goal. Such existing methods, however, do not use the clinical goal (to characterize
high-bene¿t patients) as a guide inside the construction stage. For a single outcome, recent work has shown that
this lack of linking can produce dramatically inaccurate characterizations, no matter the model.
For characterizing patients with multiple high bene¿ts, new methods must explicitly link all multiple clinical
goals (i.e. high bene¿ts in all outcomes) in the construction stage. In preparatory work, we showed that existing
methods for multiple outcomes, can miss even most of the high bene¿t patients, and we developed a preliminary
better method by establishing the missing links. This new project is motivated by our ongoing work with two
studies. The ¿rst study tested if citalopram reduces agitation in Alzheimer's patients. Since citalopram may harm
cognitive function, we set to characterize patients with high citalopram effect in (a) reducing agitation and (b)
maintaining cognitive function. The second study tests the effect of transcranial direct current stimulation on
primary progressive aphasia outcomes, with related goals. In preparatory work, we found strong evidence that
standard methods miss up to 70% of the patients with multiple high bene¿ts, compared to the new methods. For
this project we propose to fully develop methods to characterize patients who highly bene¿t on multiple outcomes.
The methods will be applied to the above studies, and can help more generally in other ADRD studies.
Aim 1. Develop methods to characterize patients who highly-bene¿t in multiple outcomes in randomized
trials. These methods are signi¿cant because they allow accurate personalized treatment choices.
Aim 2. Develop methods to ¿nd if a simpler subset of the full multiple outcomes, can have similar patient
characterization as the full outcomes. These methods are signi¿cant because they can suggest if a high-effect
on earlier outcomes is necessary before a high-effect on the later outcomes occurs.
Aim 3. Develop methods to characterize patients who highly bene¿t in multiple outcomes in observational
studies. These methods are signi¿cant when randomization is infeasible.
