Multi-Study Integer Programming Methods for Human Voltammery

点击翻译按钮获取中文摘要

Project Summary/Abstract The development of treatments for addiction requires the characterization of neural mechanisms underlying reward. Studying reward in humans requires assays that can detect changes in neurotransmitter levels with high chemical specificity. Recently, fast-scan cyclic voltammetry (FSCV) has been implemented in humans to measure dopamine with high temporal and spatial resolution. This technological achievement was enabled in large part through the novel application of machine learning methods. FSCV relies on statistical tools since FSCV records an electrochemical response which must be converted into concentration estimates via a statistical model. The validity of the scientific conclusions from human FSCV studies therefore depends heavily on the reliability of these statistical models to generate accurate dopamine concentration estimates. In human FSCV, models are fit on in vitro training sets as making in vivo training sets in humans is infeasible. Producing accurate estimates thus requires that models trained on in vitro training sets generalize to in vivo brain recordings. Combining data from multiple training sets is the standard approach human FSCV researchers have employed to improve model generalizability. This proposal extends work that shows that multi-study machine learning methods improve dopamine concentration estimates by combining training sets from different electrodes such that the resulting average signal (“cyclic voltammogram” or CV) is similar to the average CV of the electrode used in the brain. However, this approach relies on random resampling. This is problematic because the randomness limits the extent to which estimate accuracy can be improved and the slow speed of the resampling approach precludes the generation of estimates during data collection, which is critical to experiment success. This proposal details the development of methods that leverage mixed integer programming to optimally generate training sets that combine data from multiple electrodes. By generating training sets that are specifically tailored to the electrode used for brain measurements, one can vastly improve dopamine concentration estimate accuracy. The speed of the integer programming methods will enable the use of this approach during data collection. This work will include validation of the methods on in vitro data as well as on data from published in vivo and slice experiments in rodents. By applying methods to published optogenetic experiments, one can compare estimates from the proposed methods and from standard methods. The asymptotic properties of the proposed methods will be characterized analytically assuming a linear mixed effects model and empirically through application of the methods to data simulated under this model. This work will be conducted at the highly collaborative and innovative Harvard School of Public Health. The fellowship will support growth in statistical, computing and collaborative skills, and prepare the trainee for a productive career as a biostatistics professor who develops methods for neuroscience and addiction research.