There are many active research projects accessing and applying shared ADNI data. Use the search above to find specific research focuses on the active ADNI investigations. This information is requested annually as a requirement for data access.
| Principal Investigator | |
| Principal Investigator's Name: | William Goette |
| Institution: | University of Texas Southwestern Medical Center |
| Department: | Psychology |
| Country: | |
| Proposed Analysis: | The intended analyses are applications of the first-order transfer model to verbal list learning immediate recall trials. These models were developed by Igor Stepanov and Charles Abramson for mouse learning models, and they have since extended the applications of these to the learning trials of both the adult and child forms of the California Verbal Learning Test. The model is a non-linear regression composed of three parameters: attention/readiness to learn, learning rate, and learning potential. The combination of these parameters produces a dynamic learning curve with direct clinical interpretations that are of interest to neuropsychologists who interpret learning trials of list learning tests. Utilizing outpatient neuropsychological clinic data, a generalized Bayesian mixed effects non-linear model was piloted as proof-of-concept (paper presented at the 2023 International Neuropsychological Society's annual conference in San Diego). This project confirmed the clinical interpretations of the model parameters with attention being predicted by raw scores on Trail Making Test Part A and WAIS-IV Digit Span; learning rate being predicted by raw scores on processing speed measures like WAIS-IV Coding and verbal fluency; and learning potential being predicted by raw scores on language and executive functioning tasks, reflecting a likely greater memory capacity in the presence of effective existing semantic associations and test-taking strategy/adaptation. The model has since been replicated in a Danish translation of the California Verbal Learning Test and on the CERAD List Learning Test. The clinical applications of the model were tentatively explored, but the reliance on clinical data with diagnoses from various clinicians and no diagnostic validation outside of the cognitive test data limited the generalizability of the methods. ADNI data is requested to address this weakness as the determination of cognitive impairment and diagnosis is standardized and did not include the Auditory Verbal Learning Test at baseline, reducing the circularity of determining diagnostic accuracy from a test used in making the diagnosis initially. Further, ADNI's collection of medications taken, genotyping, and neuroimaging support extension of the model to greater clinical relevance. Briefly, the use of Bayesian estimation results in access to a fully described posterior distribution for the parameters of interest, and more importantly, the implied posterior predictive distribution for out-of-sample observations. Utilizing a posterior predictive distribution, it is possible to enter data for a new case and make a full prediction distribution of test performance at each trial of the list learning test. Using this distribution, it is possible to quantify a person's actual performance relative to their expected performance while including full-information regarding the uncertainty of this expectation. When a clinician is determining what diagnosis (or cause, more generally) to attribute to the case's test performance, the posterior prediction for counter-factual scenarios can be produced, and a Bayes factor quantifying the amount of evidence for the actual test performance coming from a particular scenario over another can be used to aid in diagnostic confidence. For example, if the question is whether someone's test performance represents a normal learning curve versus one likely due to amnestic MCI, then the person's data can be used to estimate two posterior predictive distributions with all other predictors being the same between the two predictions except for one where the assumption is no cognitive impairment and the other where the assumption is amnestic MCI. In most cases, this would be a simple matter of entering all the same predictor values for the two predictions but then switching a dummy variable from 0 (for no impairment) to 1 (for MCI). Returning to the specific interest in using the ADNI data with this model, the anticipated analyses intend to link commonly available data to clinicians to meaningful predictions of test performance. In the case of medication side effects, the particular interest is in examining the effects of anticholinergics, benzodiazepines, and anti-dementia medications. All of these medications are known to impact verbal learning performance, but the extent to which clinicians should adjust their expectations of what a non-impaired learning curve will look like for someone taking one (or multiple) of these medications is not well quantified. Using the methods described earlier, calculation of a learning curve with these effects included can help a clinician determine whether a person's performance on a list learning test is consistent with normal memory but medication effects or is impaired even after considering the effects of medication. Information about APOE status as well as increasingly common volumetric data from neurology are similarly clinically relevant but also lack quantified effect on clinical test performance. Beyond clinical applications, there are several theoretical extensions that are of interest. For example, how the anti-dementia medications improve or affect memory performance is important. Examination of raw scores for list learning tests, that is the sum of words correctly recalled on each learning trial, can be misleading. If an anti-dementia medication primarily improves test performance through a stimulant effect, then this results in an improved first trial but an otherwise similar learning curve. In this scenario, the medication is not improving memory itself but instead improving memory test performance by raising the person's starting point on learning tests. Using the model parameters, list learning test performance can be separated into attention, learning rate, and learning potential with the latter two being the likely most relevant targets of an anti-dementia medication. Another theoretical extension planned is an analysis of longitudinal changes in learning curves. The implications of this longitudinal extension is the ability to let clinicians detect concerning changes in test performance on re-assessments. The long-term effects of medication use can also be examined in these longitudinal models, which can be of importance for neurologists and geriatricians making decisions about medication changes in older adults. Analyses will be conducted in the R environment. The generalized Bayesian non-linear mixed effects models will be fit with the 'brms' package, which utilizes an efficient Hamiltonian Monte Carlo estimation method as implemented in the Stan language via the 'cmdstanr' package. Where the initial presentations of the first order transfer models assumed a normal likelihood function, this assumption is not reasonable for clinical settings. Previous work with the California Verbal Learning Test indicated that by the third learning trial, the model predicted scores outside the range of obtainable scores. The beta-binomial likelihood has, in my previous work, emerged as a robust alternative that has the benefit of being both bounded above and below (meaning predictions cannot be less than zero or greater than the total number of words on the list) and restricted to integer values (meaning predictions are always obtainable whole numbers). Scripts for running the models and replicating all results will also be hosted on GitHub such that any individuals who also have access to the ADNI data can fully reproduce the results. |
| Additional Investigators |
