Early Warning of Intraoperative Adverse Events via Transformer-Driven Multi-Label Learning

Xueyao Wang; Xiuding Cai; Honglin Shang; Yaoyao Zhu; Yu Yao

cs.LGMar 5, 2026

Early Warning of Intraoperative Adverse Events via Transformer-Driven Multi-Label Learning

Authors: Xueyao Wang, Xiuding Cai, Honglin Shang, Yaoyao Zhu, Yu Yao

Organizations: Chengdu Institute of Computer Application, Chinese Academy of Sciences, Chengdu 610213, China · University of Chinese Academy of Sciences, Beijing 100049, China · China Zhenhua Research Institute Co., Ltd., Guiyang 550014, China

Abstract

Early warning of intraoperative adverse events plays a vital role in reducing surgical risk and improving patient safety. While deep learning has shown promise in predicting the single adverse event, several key challenges remain: overlooking adverse event dependencies, underutilizing heterogeneous clinical data, and suffering from the class imbalance inherent in medical datasets. To address these issues, we construct the first Multi-label Adverse Events dataset (MuAE) for intraoperative adverse events prediction, covering six critical events. Next, we propose a novel Transformerbased multi-label learning framework (IAENet) that combines an improved Time-Aware Feature-wise Linear Modulation (TAFiLM) module for static covariates and dynamic variables robust fusion and complex temporal dependencies modeling. Furthermore, we introduce a Label-Constrained Reweighting Loss (LCRLoss) with co-occurrence regularization to effectively mitigate intra-event imbalance and enforce structured consistency among frequently co-occurring events. Extensive experiments demonstrate that IAENet consistently outperforms strong baselines on 5, 10, and 15-minute early warning tasks, achieving improvements of +5.05%, +2.82%, and +7.57% on average F1 score. These results highlight the potential of IAENet for supporting intelligent intraoperative decision-making in clinical practice.

Figures & tables

Figure 1: Positive sample distribution by adverse event for 5, 10, and 15 minutes prediction windows in MuAE. Percentages indicate the event occurrence rate among total samples.

Figure 2: The pipeline of Multi-adverse Events Early Warning

Figure 3: An overview of the IAENet framework for time-series in multi-label classification. Given an sample about vital sign series and static covariate in MuAE dataset $\mathbf{x}=\{\mathbf{x}_{d_{0}},...,\mathbf{x}_{d_{14}},\mathbf{x}_{s_{0}},...,\mathbf{x}_{s_{4}}\}$ , our goal is to predict whether the values in the following $\triangle$ time steps will be normal or abnormal $\mathbf{y}=\{\mathbf{y}_{1},\mathbf{y}_{2},...,\mathbf{y}_{6}\}$ . Firstly, the preprocessed dynamic variables and static covariates are respectively fused through the TAFiLM module for feature early fusion. Then, a Transformer encoder is employed to capture temporal correlations among multivariate variables. Finally, the model is trained using the proposed LCRLoss, which combines a batch-wise label frequency-weighted BCE loss with a co-occurrence constraint term based on label dependencies.

Models		IAENet		iTransformer		Crossformer		PatchTST		FEDformer		Autoformer		Informer		DLinear		Non-stationary		TFT		SegRNN
	$\triangle$	F1	AUC	F1	AUC	F1	AUC	F1	AUC	F1	AUC	F1	AUC	F1	AUC	F1	AUC	F1	AUC	F1	AUC	F1	AUC
Hypotension	5	51.45	74.83	45.64	65.76	48.32	67.29	24.30	57.58	45.90	66.69	3.02	50.69	47.65	67.96	47.65	67.69	48.37	68.66	36.71	62.59	29.23	59.19
	10	36.21	65.85	16.91	54.79	19.14	55.50	10.58	52.80	11.59	53.11	1.07	50.10	25.16	58.35	17.81	55.20	23.10	57.40	19.09	55.54	14.47	54.55
	15	28.54	62.08	9.79	52.57	15.24	54.28	5.45	51.29	10.13	52.67	0.03	49.95	11.17	53.02	8.01	52.03	16.10	54.84	16.81	55.31	10.17	52.68
Low DoA	5	61.45	76.69	54.03	69.49	37.70	66.77	2.28	50.43	51.42	68.08	6.11	50.31	49.53	67.19	49.53	67.19	47.92	66.19	6.18	51.06	54.65	70.35
	10	56.54	73.42	49.80	67.35	43.18	63.80	0.95	50.17	43.85	64.14	11.62	50.57	37.89	61.17	43.77	64.09	31.35	58.56	7.08	51.21	43.59	64.01

Table 1: Classification Accuracy results across all tasks and methods.

\triangle

denotes the forecasting horizon time steps. All metrics are in %. The best results are in bold font and the second best are underlined .

Loss	ACC	PRE	REC	F1	AUC	HM
LCRLoss	91.72	61.77	69.39	65.36	80.88	0.0828
WBCE	92.17	68.22	56.99	62.10	71.12	0.0783
BCE	92.69	71.70	57.83	64.03	74.21	0.0731
FL	92.73	73.76	54.91	62.96	73.75	0.0727
ASL	91.22	58.98	72.08	64.87	79.86	0.0878
PolyLoss	92.66	70.90	58.91	64.35	75.07	0.0734

Table 2: Performance of IAENet under different losses.

Module	ACC	PRE	REC	F1	AUC	HM
Informer	91.28	65.74	46.92	54.76	71.54	0.0872
+ LCRLoss	90.24	55.87	62.89	59.17	79.36	0.0976
+ FiLM	92.05	70.10	51.14	59.13	70.62	0.0795
+ TAFiLM	91.95	67.89	53.96	60.13	71.34	0.0805
Crossformer	92.39	73.03	51.36	60.31	71.56	0.0761
+ LCRLoss	90.80	57.88	66.80	62.02	81.17	0.0920

Table 3: Ablation studies on the components of IAENet.

Factor	Setting	F1	AUC
Weighting (G / L)	none	63.97	74.30
	inverse	40.55 / 60.82	78.80 / 85.35
	log_inverse	24.40 / 51.02	60.23 / 83.94
	sqrt_inverse	59.51 / 65.36	83.68 / 80.88
	cubic_inverse	63.91 / 64.26	82.33 / 79.44
Batch Size	64	65.21	81.35

Table 4: Ablation studies on reweighting strategies, batch size, and

\lambda

of LCRLoss.

Figure 4: Derivatives of the loss functions. The X-axis denotes the logit of positive labels, and the Y-axis is the corresponding gradients.

Figure 5: Co-occurrence matrix of adverse events

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 1: Non-Missing rate for variables in the MuAE dataset

	Training dataset	Validation dataset	Test dataset	P-value	Test
	(n = 610, 70%)	(n = 88, 10%)	(n = 175, 20%)	P-value	Test
Age, years $\pm$ SD	59.59 $\pm$ 14.15	61.22 $\pm$ 15.48	60.88 $\pm$ 12.98	0.502	Kruskal-Wallis
Sex, n(%)				0.776	Chi-squared
Female	257 (42.1)	38 (43.2)	79 (45.1)
Male	353 (57.9)	50 (56.8)	96 (54.9)
Weight	62.94 $\pm$ 11.99	62.87 $\pm$ 11.76	61.53 $\pm$ 10.57	0.379	Kruskal-Wallis

Appendix

Table 1: Baseline characteristics of the dataset

Device/Variable	Description	Covariate
Age	Age of the patient	$\checkmark$
Sex	Gender of the patient	$\checkmark$
Weight	Body weight of the patient	$\checkmark$
Height	Height of the patient	$\checkmark$
ASA	ASA classification	$\checkmark$
Orchestra/PPF20_VOL	Propofol infusion volume

Appendix

Table 2: Summary of Variables Selected from the MuAE Dataset

Figure 2: Gradient Analysis for Negative Sample

Figure 3: Gradient Analysis for Positive Sample

Clinical Risk Prediction

Clinical risk prediction estimates patients’ future health outcomes from medical records, physiological signals, imaging, and other health data.

See all 217 papers

Aug 30, 2026

INTERVenE: Temporal-Abstraction-Interval Based Transformers for Short-Horizon Medical Event Prediction

Electronic Health Record (EHR) prediction models in the intensive care unit must learn from sparse and irregular measurements while preserving the clinical meaning of time and supporting transparent decision-making. We present INTERVenE, a family of Transformer architectures whose input is an interval-based, knowledge-based temporal abstraction (KBTA), a token stream of named clinical concepts (states, trends, events, contexts) drawn from a curated medical ontology, rather than an unnamed bin index or a raw measurement triplet. This naming layer is what we ask KBTA to do: it makes the model's per-token attributions resolve to clinical concepts by construction. INTERVenE offers two complementary variants: an auto-regressive decoder that generates future abstraction trajectories with a per-step risk readout (localizing when and after which events risk rises), and a bidirectional encoder for single-pass joint risk and time-to-event prediction. On 57,078 MIMIC-IV admissions, INTERVenE-Enc outperformed GRU-D, STraTS, and KarmaLego, reaching a support-weighted AUPRC of 0.672 (

+0.046

over the strongest neural baseline, with non-overlapping 95% bootstrap CIs), AUROC of 0.901, and length-of-stay MAE of 44.4,h. INTERVenE-Ar complements these predictions with generated abstraction trajectories and evolving risk scores, achieving AUROC of 0.854 and AUPRC of 0.588 under the harder generative readout. An alternative-discretization ablation preserved predictive performance, supporting KBTA's role as a clinically named attribution substrate rather than a prerequisite for accuracy.

Shahar Oded, Yuval Shahar

The Stein Faculty of Computer and Information Science, Ben-Gurion University of the Negev, Beer-Sheva, Israel

Event Time Prediction Interpretable ML Clinical Outcome Prediction Electronic Health Records Time Series Transformers

Ask PDF

Jun 22, 2026

DBT-Bleed: Dual-Branch Temporal Modeling with Key-Frame Selection for Surgical Bleeding Detection

Intraoperative Adverse Events (IAEs) detection is critical for improving surgical safety, with bleeding being among the most frequent events across many surgery types. Existing methods struggle to distinguish bleeding IAE from visually similar residual blood due to limited temporal reasoning. Moreover, modeling long surgical videos while preserving fine-grained temporal dynamics remains computationally challenging. We propose DBT-Bleed, a dual-branch multi-scale temporal modeling framework disentangling bleeding and normal representations using layer-wise temporal adapters for short- and long-term bleeding progression. To efficiently process long surgical videos without sacrificing fine-grained temporal information, we introduce HiRED, a Hierarchical Entropy-Driven frame selection strategy that retains temporally informative segments while removing redundancy. Experiments on the MultiBypass dataset demonstrate gains of 6.53% in F1, 5.62% in Recall and 9% in MCC values for bleeding IAE detection, consistently outperforming video-level baselines. Additionally, we evaluate cross-procedure generalization on a newly curated dataset from a different surgical procedure type, where DBT-Bleed demonstrates robust transferability by achieving gain of 6% in F1 and 8% in MCC under zero-shot setting. To support this evaluation, we introduce EndoPit-IAE, an Endonasal Pituitary Surgery dataset annotated for IAEs, representing the first IAE-annotated dataset in neurosurgery. Code will be made publicly available upon acceptance.

Sudhanshu Mishra, Jialang Xu, Jensen Ang +3

Department of Electrical and Computer Engineering, National University of Singapore, Singapore · UCL Hawkes Institute, Department of Medical Physics and Biomedical Engineering, University College London, London, United Kingdom · Department of Neurosurgery, National Neuroscience Institute, Singapore +1

Surgical Video Understanding Video Frame Selection Temporal Video Understanding Anomaly Detection

Ask PDF

Apr 2, 2026

MATA-Former & SIICU: Semantic Aware Temporal Alignment for High-Fidelity ICU Risk Prediction

Forecasting evolving clinical risks relies on intrinsic pathological dependencies rather than mere chronological proximity, yet current methods struggle with coarse binary supervision and physical timestamps. To align predictive modeling with clinical logic, we propose the Medical-semantics Aware Time-ALiBi Transformer (MATA-Former), utilizing event semantics to dynamically parameterize attention weights to prioritize causal validity over time lags. Furthermore, we introduce Plateau-Gaussian Soft Labeling (PSL), reformulating binary classification into continuous multi-horizon regression for full-trajectory risk modeling. Evaluated on SIICU -- a newly constructed dataset featuring over 506k events with rigorous expert-verified, fine-grained annotations -- and the MIMIC-IV dataset, our framework demonstrates superior efficacy and robust generalization in capturing risks from text-intensive, irregular clinical time series.

Zhichong Zheng, Xiaohang Nie, Xueqi Wang +3

Tongji University · Shanghai Innovation Institute · Harbin Institute of Technology +1

Clinical Risk Prediction Longitudinal EHR Modeling Irregular Time-Series Modeling Temporal Attention

Ask PDF

Abstract

Clinical Risk Prediction

Explore similar work

INTERVenE: Temporal-Abstraction-Interval Based Transformers for Short-Horizon Medical Event Prediction

DBT-Bleed: Dual-Branch Temporal Modeling with Key-Frame Selection for Surgical Bleeding Detection

MATA-Former & SIICU: Semantic Aware Temporal Alignment for High-Fidelity ICU Risk Prediction