Robust missing data reconstruction in schizophrenia using tracking-removed autoencoder with fuzzy confidence integration.

Neural network models for outcome prediction play a pivotal role in neurological disease research, particularly for baseline risk assessment. Schizophrenia, a complex and relatively rare neuropsychiatric disorder, presents significant diagnostic challenges due to its heterogeneous symptomatology and reliance on subjective clinical evaluations. One of the primary barriers to accurate modelling in Schizophrenia research is the prevalence of incomplete clinical data, often stemming from inconsistent reporting, participant dropout, and variability in assessment protocols. To address these limitations, an advanced deep learning framework has been developed that integrates the Tracking-Removed Autoencoder (TRAE) with Multi-View Progressive Training (MVPT). This architecture is specifically designed for predicting missing features within schizophrenia datasets. The model treats missing data as learnable information rather than simply marking it as absent. By restructuring how inputs are processed, it incorporates these gaps directly into training, allowing the system to detect meaningful patterns and remain robust even when clinical records are incomplete. Fuzzy confidence measures assess how reliable each imputed feature is by determining its membership in clinically meaningful categories, or fuzzy sets. This methodology, grounded in fuzzy set theory, can generate linguistic confidence measures, converting fuzzy memberships into interpretable descriptors that convey meaningful information to clinicians and researchers. The linguistic framework captures uncertainty and complexity while facilitating the integration of higher-level clinical knowledge. Within the TRAE + MVPT framework, this means the model can not only fill in missing data but also indicate how much confidence clinicians and researchers should place in each reconstructed value. Such quantification is particularly important in schizophrenia, where overlapping symptoms and diagnostic ambiguity often reduce clarity. This novel framework strengthens psychiatric data modelling by improving both the interpretability and reliability of reconstructed clinical features, directly addressing challenges of diagnostic uncertainty. It supports more transparent decision-making and clearer symptom tracking, while in research it provides a rigorous way to evaluate both model reliability and imputation quality. Although neural network-based imputation has been used in other areas of medicine, the integration of Multi-View Progressive Training (MVPT) with fuzzy confidence measures has not yet been applied to schizophrenia datasets.