# Multimodal Autoencoder–Based Anomaly Detection Reveals Clinical–Radiologic Heterogeneity in Pulmonary Fibrosis

## TL;DR

This study uses AI to detect unusual patterns in lung disease patients by combining imaging and clinical data, revealing hidden variability beyond traditional severity measures.

## Contribution

A novel multimodal AI framework using a variational autoencoder for unsupervised anomaly detection in pulmonary fibrosis.

## Key findings

- Anomaly scores identified 17.1% of patients with atypical clinical–radiologic profiles across all severity categories.
- Anomaly scores showed weak correlation with traditional severity markers like DLCO and FEV1.
- Highly anomalous patients often had mismatched clinical and radiologic features.

## Abstract

Background: Pulmonary fibrosis (PF) and post-infectious fibrotic lung disease are characterized by marked heterogeneity in radiologic patterns, physiologic impairment, and clinical presentation. Conventional analytic approaches often fail to capture non-linear and multimodal relationships between structural imaging findings and functional limitation. Integrating imaging-derived representations with clinical and functional data using artificial intelligence (AI) may provide a more comprehensive characterization of disease heterogeneity. Objectives: The objective of this study was to develop and evaluate a multimodal AI framework combining imaging-derived embeddings and structured clinical data to identify atypical clinical–radiologic profiles in patients with pulmonary fibrosis using unsupervised anomaly detection. Methods: A retrospective cohort of 41 patients with radiologically confirmed pulmonary fibrosis or post-infectious fibrotic lung disease was analyzed. Deep imaging embeddings were extracted from baseline thoracic CT examinations using a pretrained convolutional neural network and integrated with standardized clinical and functional variables. A multimodal variational autoencoder (VAE) was trained in an unsupervised manner to learn the distribution of typical patient profiles. Patient-specific anomaly scores were derived from reconstruction error plus latent regularization (β·KL divergence). Associations between anomaly scores, disease severity, and clinical markers were assessed using Spearman rank correlation. Results: Anomaly scores were right-skewed (median 26.91, IQR 22.87–32.11; range 19.75–46.18). Patients above the 85th percentile (anomaly score ≥ 33.85) comprised 7/41 (17.1%) of the cohort and occurred across all clinician-assigned severity categories (mild 3, moderate 1, severe 3). Anomaly scores overlapped substantially across severity groups, with similar medians (mild 26.47, moderate 28.55, severe 28.23). Correlations with conventional severity markers were weak and non-significant, including DLCO (% predicted; ρ = −0.25, p = 0.115) and FEV1 (% predicted; ρ = −0.22, p = 0.165), a pattern consistent with anomaly scores reflecting multimodal deviation rather than severity alone, while acknowledging the exploratory nature of the analysis. Highly anomalous patients frequently exhibited discordant clinical–radiologic profiles, including preserved functional capacity despite marked imaging-derived deviation or disproportionate physiological impairment relative to imaging patterns. Conclusions: This proof-of-concept study demonstrates that multimodal VAE-based anomaly detection integrating imaging-derived embeddings with clinical data can quantify clinical–radiologic heterogeneity in pulmonary fibrosis beyond conventional severity stratification. Unsupervised anomaly detection provides a complementary framework for identifying atypical multimodal profiles and supporting individualized phenotyping and hypothesis generation in fibrotic lung disease. Given the modest cohort size, these findings should be interpreted as illustrative and hypothesis-generating rather than generalizable.