A transfer learning-based multimodal model for early prediction of 90-day respiratory failure in
Lihui Guo1, Yaning Yao1, Qirui Wu2
1Department of Ultrasound, The First Hospital of Lanzhou University, Lanzhou, Gansu, China.
Background:
Dermatomyositis-associated interstitial lung disease (DM-ILD) is a life-threatening condition that often leads to respiratory failure, poor prognosis, and high mortality, especially in patients positive for anti-melanoma differentiation-associated gene 5 (anti-MDA5) antibodies. Early identification of MDA5 positivity is crucial for timely intervention, yet routine testing is not consistently available across different regions. In some centers, the test is omitted or sent to external laboratories, which can delay recognition of disease severity and postpone appropriate treatment. Because risk assessment based on admission data remains limited, we focused on 90-day respiratory failure as an early and clinically relevant outcome, collecting patients' clinical information within the first 48 hours of admission while excluding anti-MDA5 results. Using these data, we developed and internally tested a multimodal model to predict 90-day respiratory failure in patients with DM-ILD.
Methods:
In this exploratory single-center retrospective study, 124 adult patients with DM-ILD were analyzed using a 7:3 train-test split, with five-fold cross-validation used within the training set. Baseline predictors included demographic characteristics, clinical features, laboratory results, pulmonary function indices, and latent computed tomography (CT) features extracted from chest CT scans using a pre-trained model, all collected within 48 hours of admission. Multiple dimensionality-reduction, modeling, and fusion strategies were compared to identify the optimal framework. We evaluated model performance using the area under the receiver operating characteristic curve (AUC) and assessed interpretability with SHapley Additive exPlanations (SHAP).
Results:
Among the prespecified candidate models, the early-fusion random forest model incorporating principal component analysis (PCA) showed the best discriminative performance in the held-out test set, with an AUC of 0.967 and a PR-AUC of 0.879. SHAP analysis showed that arthritis, pulmonary function indices, laboratory markers, and several latent CT features were among the most influential predictors.
Conclusions:
The admission-based multimodal model demonstrated encouraging performance in internal testing and may aid early stratification of 90-day respiratory failure risk in patients with DM-ILD, particularly when anti-MDA5 antibody results are unavailable or delayed.
