High-density retinal signal deciphering in support of diagnosis in psychiatric disorders: A new paradigm

Abstract

Background: The search for biomarkers in psychiatry started decades ago with major reported milestones in the development of computational psychiatry and precision medicine. Prediction models have been suggested based upon components extracted from retinal signals and analyses of electroretinogram (ERG), with the objective of providing prediction metrics to support diagnoses. However, conventional ERG parameters lack detailed information to appropriately decipher retinal signals and extract specific descriptors that best describe such pathologies. Methods: We developed the concept of high-density retinal signal, with the specific target of modeling mathematical domains of information gathered from retinal signals and related clinical information. An interim analysis has been conducted in the framework of a multicenter clinical study, aiming to develop prediction models that differentiate between two major psychiatric disorders, schizophrenia and type 1 bipolar disorder. In order to select the best predictors within the entirety of the signals, and the full extent of the available information, in addition to using conventional ERG parameters, two new approaches for extracting non-conventional retinal signal descriptors were implemented. Results: Mappings of predictors demonstrate that meaningful classifiers are located in areas of the retinal signal that have never been investigated before, allowing a multiplicity of biomarkers to be extracted, all well scattered within the entire volume of information, as opposed to the conventional ERG components which are very sparce and discrete. RSPA prediction models minimum accuracy was 79% and maximum 99% for training, and 68% and 90%, respectively, for testing, depending upon the model used, as compared to 73% and 87% for training, and 55% and 61% for the testing dataset with the prediction models using conventional ERG parameters alone. The prediction models with the highest testing performance were found using Ridge logistic regression with either photopic MA, ARMA or signal density polynomial coefficients predictors. Meaningful testing performances were also obtained with logistic regression (90%), neural network (88%) and SVM (86%) analysis methods. These results demonstrate that using only conventional ERG parameters is a very limited approach in prediction model development, because it excludes most of the retinal signal where many specific details are the most performing classifiers. Conclusion: Our findings support the concept of high-density retinal signal and its purpose, while many research groups attempt to decipher retinal signals to differentiate between complex pathologies, such as psychiatric disorders, to select biosignatures as objective evidence for their diagnoses.