NeoTriFuse: Reliability-Aware Multimodal Fusion under Missingness Heterogeneity for Neonatal Mortality Risk Prediction

Jiyuan Tian; Qincheng Shen; Ye Lin; Yu Gao; Haohui Lu

cs.LGAug 26, 2026

NeoTriFuse: Reliability-Aware Multimodal Fusion under Missingness Heterogeneity for Neonatal Mortality Risk Prediction

Authors: Jiyuan Tian, Qincheng Shen, Ye Lin, Yu Gao, Haohui Lu

Organizations: Discipline of Business Analytics, The University of Sydney, Sydney, Australia · Molly Wardaguga Institute for First Nations Birth Rights, Charles Darwin University, Darwin, Australia

Abstract

Neonatal mortality risk prediction from bedside monitoring data remains challenging due to extreme class imbalance, heterogeneous clinical risk factors, multi-scale temporal dynamics, and substantial missingness. We propose NeoTriFuse, a reliability-aware multimodal fusion framework for missingness-heterogeneous neonatal monitoring data. Unlike conventional multimodal approaches that treat missingness primarily as a preprocessing issue, NeoTriFuse models missingness as an explicit reliability signal that dynamically modulates modality contributions during fusion. The framework integrates static perinatal variables, local-global temporal encoders, and patient-level statistical summaries through reliability-guided gating mechanisms, while jointly optimizing mortality prediction and an auxiliary length-of-stay objective. NeoTriFuse achieves competitive performance, with an F1 score of 0.6736 +/- 0.0216 and an AUROC of 0.9454 +/- 0.0056. Ablation studies indicate that the local-global temporal architecture and patient-level summary branch contribute most substantially to predictive performance, while reliability-aware gating provides additional improvements on threshold-dependent metrics under heterogeneous observation completeness. Sensitivity analyses further suggest stable performance across nearby hyperparameter settings. Overall, the findings support reliability-aware multimodal fusion as a practical approach for neonatal mortality prediction under realistic clinical missingness conditions.

Figures & tables

Figure 1: Overview of NeoTriFuse for reliability-aware multimodal fusion under heterogeneous missingness. The model integrates temporal HR and SpO 2 sequences, static clinical variables, patient-level summary statistics, and a reliability vector derived from missingness and observation coverage. Local and global temporal encoders generate $\mathbf{h}_{temporal}$ , while reliability-guided gates modulate temporal, static, and summary representations before fusion. The fused representation is used for mortality prediction with auxiliary length-of-stay prediction.

Model	F1	Precision	Recall	AUROC
TIM-MSFL	0.6570 $\pm$ 0.0126	0.7780 $\pm$ 0.0370	0.5888 $\pm$ 0.0170	0.9292 $\pm$ 0.0014
GRU	0.6446 $\pm$ 0.0117	0.7246 $\pm$ 0.0421	0.6147 $\pm$ 0.0245	0.9450 $\pm$ 0.0019
CNN	0.6348 $\pm$ 0.0120	0.7445 $\pm$ 0.0310	0.5916 $\pm$ 0.0183	0.9317 $\pm$ 0.0042
Transformer	0.6201 $\pm$ 0.0164	0.6397 $\pm$ 0.0314	0.6430 $\pm$ 0.0211	0.9316 $\pm$ 0.0037
TNformer-MP	0.6155 $\pm$ 0.0202	0.6314 $\pm$ 0.0357	0.6273 $\pm$ 0.0205	0.9303 $\pm$ 0.0039
GRU-D	0.5690 $\pm$ 0.0223	0.6442 $\pm$ 0.0438	0.5739 $\pm$ 0.0234	0.9250 $\pm$ 0.0030

Table 1: Comparison with traditional, neural, and missingness-aware baselines. Values are mean

\pm

std over 10 seeds.

Figure 2: Heatmap of classification performance across baselines and the proposed model.

Variant	F1	Precision	Recall	AUROC
No Local CNN	0.6632 $\pm$ 0.0095	0.6961 $\pm$ 0.0319	0.6824 $\pm$ 0.0210	0.9450 $\pm$ 0.0038
No Global Transformer	0.6412 $\pm$ 0.0171	0.6633 $\pm$ 0.0381	0.6750 $\pm$ 0.0267	0.9292 $\pm$ 0.0078
No Tabular Summary	0.6470 $\pm$ 0.0115	0.6949 $\pm$ 0.0332	0.6633 $\pm$ 0.0232	0.9420 $\pm$ 0.0080
Concat Fusion	0.6679 $\pm$ 0.0090	0.7427 $\pm$ 0.0309	0.6539 $\pm$ 0.0213	0.9436 $\pm$ 0.0048
No Reliability Vector	0.6685 $\pm$ 0.0127	0.7157 $\pm$ 0.0513	0.6663 $\pm$ 0.0335	0.9473 $\pm$ 0.0052
No LOS Auxiliary	0.6641 $\pm$ 0.0156	0.7076 $\pm$ 0.0588	0.6735 $\pm$ 0.0309	0.9372 $\pm$ 0.0045

Table 2: Ablation study of the proposed model.

Config	F1	Precision	Recall	AUROC
hd128_los0.3	0.6690 $\pm$ 0.0162	0.7144 $\pm$ 0.0370	0.6778 $\pm$ 0.0233	0.9419 $\pm$ 0.0037
hd96_los0.2	0.6662 $\pm$ 0.0164	0.6999 $\pm$ 0.0403	0.6824 $\pm$ 0.0252	0.9439 $\pm$ 0.0060
hd128_los0.2	0.6658 $\pm$ 0.0144	0.6945 $\pm$ 0.0414	0.6785 $\pm$ 0.0329	0.9457 $\pm$ 0.0044
hd64_los0.3	0.6653 $\pm$ 0.0142	0.7218 $\pm$ 0.0573	0.6653 $\pm$ 0.0424	0.9409 $\pm$ 0.0058
hd96_los0.1	0.6629 $\pm$ 0.0204	0.7060 $\pm$ 0.0407	0.6711 $\pm$ 0.0307	0.9419 $\pm$ 0.0079
hd128_los0.1	0.6617 $\pm$ 0.0128	0.7112 $\pm$ 0.0336	0.6667 $\pm$ 0.0237	0.9443 $\pm$ 0.0049

Table 3: Ten-seed sensitivity analysis of hidden dimension and LOS loss weight.

Figure 3: Sensitivity of F1 and AUROC across hidden dimension and LOS loss weight settings.

Model	Cohort and input data	Method	Results	Comment
Feng et al. [ 4 ]	285 low birth weight preterm infants; clinical variables and vital signs	Deep learning	ACC: 0.888 AUROC: 0.897 Recall: 0.780	Strong temporal model; small cohort
Niestroy et al. [ 18 ]	5,957 NICU infants; daily 10-min HR and SpO 2 segments	Time-series features with logistic regression	AUROC: 0.828	Effective feature engineering
Li et al. [ 12 ]	459 MIMIC-III infants, 23–29 weeks; admission features	Machine learning models	AUROC: 0.91 F1: 0.67	Admission-based prediction
Sullivan et al. [ 25 ]	PAS challenge data; demographics and HR/SpO 2 samples	Machine learning models	AUROC: 0.818	Strong PAS baseline
NeoTriFuse	PAS challenge data; multimodal inputs including static variables, HR/SpO 2 , and masks	Reliability-aware multimodal fusion	AUROC: 0.9454 F1: 0.6736	Missingness as reliability signal

Table 4: Comparison with related neonatal mortality and NICU survival prediction studies.

Reliability-Aware Multimodal Fusion

See all 23 papers

Evidential Fusion Network for Multimodal Survival Prediction under Missing Modalities
Jun 18, 2026Yucheng Xing, Hailan Mo, Zi Wang +2Missing-Modality Learning Uncertainty Quantification
National University of Singapore, Singapore · Imperial College London, London, UK
Reliability-Aware Probabilistic Fusion for Multimodal Learning with Incomplete Evidence
Oct 21, 2026Patitapaban Palo, Abhisek Tiwari, Mattia Carletti +3Reliability-Aware Multimodal Fusion Uncertainty-Aware Multimodal Fusion
Institute of Biomedical Engineering, Department of Engineering Science, University of Oxford, UK
TRACE: Tracking Multimodal Clinical Trajectories for Inpatient Risk Prediction
Oct 19, 2026Minjun Kim, Jong Hak MoonClinical Outcome Prediction Clinical Risk Prediction
Yeji X

Jun 18, 2026J/K move · Enter open · S save

Evidential Fusion Network for Multimodal Survival Prediction under Missing Modalities

Yucheng Xing, Hailan Mo, Zi Wang, Ling Huang+1

National University of Singapore, Singapore · Imperial College London, London, UK

Missing-Modality Learning Uncertainty Quantification Cancer Survival Prediction Survival Analysis Dempster-Shafer Theory Uncertainty-Aware Multimodal Fusion+3

Open paper AskPDF arXiv

Recent multimodal survival prediction models have demonstrated strong predictive performance by leveraging complementary information across modalities. However, such models generally assume data completeness and exhibit limited robustness toward missing modalities, which are frequently encountered in real-world clinical settings. We propose the Evidential Missing Modality Survival Fusion (EMMS) model for multimodal survival prediction under missing modalities. EMMS offers a straightforward, computationally effective approach to survival analysis without requiring a generative phase for missing data. By employing Dempster-Shafer theory and Gaussian Random Fuzzy Numbers for multimodal decision fusion, it considers both aleatoric and epistemic uncertainty alongside modality reliability for fusion. Moreover, the model treats missing modalities as vacuous evidence, preventing interference with available inputs and naturally reflecting increased uncertainty and calibrated predictions. Extensive experiments on four cancer datasets demonstrate state-of-the-art performance while providing calibrated and interpretable uncertainty estimates under incomplete multimodal observations, without introducing additional computational overhead.

Abstract

Reliability-Aware Multimodal Fusion

Explore similar work

Evidential Fusion Network for Multimodal Survival Prediction under Missing Modalities

Reliability-Aware Probabilistic Fusion for Multimodal Learning with Incomplete Evidence

TRACE: Tracking Multimodal Clinical Trajectories for Inpatient Risk Prediction