This page was translated automatically from Turkish. Read the original

Bu sayfa, 26. Dünya Psikiyatri Kongresi'nde sunulan çalışmanın yapay zekâ desteğiyle hazırlanmış Türkçe özetidir. Çalışmanın sahibi sunan yazarlardır; atıf için orijinal kaynağa başvurun.

EV0310

The geometry of trauma: mapping latent PTSD symptom profiles with sparse autoencoders

THE GEOMETRY OF TRAUMA: MAPPING LATENT PTSD SYMPTOM PROFILES VIA SPARSE AUTOENCODERS

A six-dimensional symptom map learned by a sparse autoencoder grouped patients with identical PCL-5 total scores more consistently and reconstructed their profiles more accurately than linear methods.

The aim was to test whether a symptom map learned from the same 20 PCL-5 items could reconstruct patients’ profiles more accurately than linear methods and keep patients with identical total scores within a single coherent region. The rationale was that although PTSD severity is reported as a single PCL-5 total score, this score sums 20 different symptoms, and two patients who both score 37 may have entirely different clinical presentations. The sample comprised 3,311 PCL-5 assessments; the data were split 80/20 (test n = 663, 76 score groups). A sparse autoencoder (from 20 dimensions to 6 and back to 20), principal component analysis (PCA), and factor analysis were compared; outcome measures were reconstruction error, explained variance, and within-score dispersion (WSD), defined as the mean distance between patients sharing the same total score. Robustness was checked using standardized maps, 1,000 bootstrap resamples, and 5 random initializations; six dimensions were selected to align with established PTSD symptom cluster models. The sparse autoencoder performed best (MSE 0.264, explained variance 73.6%, WSD 2.48); the corresponding values were 0.286, 72.1%, and 2.65 for PCA, and 0.338, 66.2%, and 2.61 for factor analysis. As an example, on the z5 arousal dimension, difficulty concentrating and hypervigilance had positive loadings of +0.28 and +0.23, respectively, while startle response, irritability and anger, and reckless behavior had negative loadings of −0.21, −0.21, and −0.38, respectively; this distinguishes inwardly directed from outwardly directed arousal within PCL-5 Cluster E. The authors propose that patients should be characterized by their location in symptom space as well as their total score, that changes during follow-up could be interpreted as movement within this stable space, and that the arousal dimension could help guide treatment matching by distinguishing a tense, hypervigilant presentation from an angry, impulsive one. Limitations include the cross-sectional design, the modest effect size and lack of validation against clinical outcomes to date, the absence of comparisons with IRT or bifactor models, and moderate stability (r = 0.59).

Tüm WCP 2026 özetleri