ToxiNet: A Multimodal Deep Learning Framework for Predicting Clinical Outcomes in Acute Drug Overdose Using Electronic Health Records
International Journal of Medical Toxicology and Forensic Medicine,
Vol. 16 (2026),
1 Dey 2026
,
Page 1-11
https://doi.org/10.22037/ijmtfm.v16.53708
Abstract
Background: Acute drug overdose is a leading cause of preventable emergency department (ED) visits and in-hospital mortality worldwide, and clinicians must often make rapid triage, monitoring, and treatment-escalation decisions from fragmented, multimodal data, structured vitals and labs, free-text clinical notes, toxicology assays, and cardiac telemetry, under significant time pressure. Existing prediction tools typically rely on a single data modality and on static risk scores, limiting their sensitivity to the fast-evolving physiological trajectory of an overdose patient.
Methods: ToxiNet introduces a multimodal deep learning framework that combines electronic health record time series, clinical text, toxicology panels, and ECG waveforms through modality-specific encoders and cross-modal attention fusion to jointly predict thirty-day mortality, ICU admission, length of stay, major complications, and clinical recommendations.
Results: In a controlled, held-out evaluation, ToxiNet achieved an AUC-ROC of 0.93 for the composite mortality/ICU-admission outcome, compared with 0.86 for an LSTM baseline using structured EHR data alone, 0.81 for a random forest baseline, and 0.77 for logistic regression. An incremental modality-ablation analysis showed AUC rising from 0.83 (structured EHR only) to 0.93 as clinical notes, toxicology panel, and ECG modalities were progressively added, with the largest single-modality contribution coming from the toxicology panel.
Conclusion: These results, obtained in a controlled and currently simulation-supported evaluation, suggest that multimodal fusion of structured, textual, toxicological, and cardiac signal data can improve early risk stratification in acute overdose beyond what any single modality supports. As detailed in the Methods and Limitations sections, prospective, multi-site validation on real, IRB-approved clinical data together with formal statistical testing and toxicologist-in-the-loop evaluation is required before ToxiNet’s outputs can support real clinical decisions.
- Acute Drug Overdose, Multimodal Deep Learning, Electronic Health Records, Clinical Decision Support, Medical Toxicology, Cross-Modal Attention
How to Cite
References
[1] Dong X, et al. Predicting opioid overdose risk of patients with opioid prescriptions using electronic health records based on temporal deep learning. J Biomed Inform. 2021;116:103725. [DOI: 10.1016/j.jbi.2021.103725]
[2] Dong X, et al. Machine learning based opioid overdose prediction using electronic health records. In: Proc AMIA Annu Symp. 2019. p. 389-98.
[3] Ramírez Medina CR, et al. A systematic review of machine learning applications in predicting opioid associated adverse events. npj Digit Med. 2025;8:30. [DOI: 10.1038/s41746-024-01312-4]
[4] Gellad WF, et al. Development and validation of an overdose risk prediction tool using prescription drug monitoring program data. Drug Alcohol Depend. 2023;246:109856. [DOI: 10.1016/j.drugalcdep.2023.109856]
[5] Lo-Ciganic WH, et al. Integrating human services and criminal justice data with claims data to predict risk of opioid overdose among Medicaid beneficiaries: a machine-learning approach. PLoS ONE. 2021;16(3):e0248360. [DOI: 10.1371/journal.pone.0248360]
[6] A deep learning method to detect opioid prescription and opioid use disorder from electronic health records. 2023.
[7] Zandi M, Nasimzade S, Pourhossein B, Fazeli M. A Snapshot of Different Types of Under-Research Vaccines A Gainst Covid 1 9: A Review. 2020.
[8] Alsentzer E, et al. Publicly available clinical BERT embeddings. In: Proc 2nd Clin Nat Lang Process Workshop. 2019. p. 72–78.
[9] Lundberg SM, Lee SI. A unified approach to interpreting model predictions. In: Proc Adv Neural Inf Process Syst (NeurIPS). 2017;30.
[10] Johnson AEW, et al. MIMIC-III, a freely accessible critical care database. Sci Data. 2016;3:160035. [DOI: 10.1038/sdata.2016.35]
[11] Naresh P, Namratha P, Kavitha T, Chaganti S, Elicherla SLR, Gupta KG. Utilizing machine learning for the identification of chronic heart failure (CHF) from heart pulsations. In: 2024 4th Int Conf Ubiquitous Comput Intell Inf Syst (ICUIS). 2024. p. 1037–42. [DOI: 10.1109/ICUIS64676.2024.10866468]
[12] Hannun AY, et al. Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network. Nat Med. 2019;25:65–69. [DOI: 10.1038/s41591-018-0268-3]
[13] Che Z, Purushotham S, Cho K, Sontag D, Liu Y. Recurrent neural networks for multivariate time series with missing values. Sci Rep. 2018;8:6085. [DOI: 10.1038/s41598-018-24271-9]
[14] Rumack BH, Matthew H. Acetaminophen poisoning and toxicity. Pediatrics. 1975;55(6):871-6.
[15] Wang C, Yang X, Sun M, Gu Y, Niu J, Zhang W. Multimodal fusion network for ICU patient outcome prediction. Neural Netw. 2024;180.
[16] Aruna T, Naresh P, Kumar BA, Prakash BK, Mohan KM, Reddy PM. Analyzing and detecting digital counterfeit images using DenseNet, ResNet and CNN. In: 2024 8th Int Conf Inventive Syst Control (ICISC). 2024. p. 248–52. [DOI: 10.1109/ICISC62624.2024.00049]
[17] Abuhamad H, Zainudin S, Abu Bakar A. Integrative multimodal hybrid data fusion for mortality prediction. Sci Rep. 2026;16:5803.
[18] Sadanandan B. Multimodal deep learning for early prediction of patient deterioration in the ICU: integrating time-series EHR data with clinical notes. arXiv:2603.14719. 2026.
[19] Alcaraz JML, et al. A multimodal deep learning framework for predicting ICU deterioration: integrating ECG waveforms with clinical data and clinician benchmarking. arXiv:2601.06645. 2026.
[20] Zhang Y, et al. An empirical study of using radiology reports and images to improve intensive care unit mortality prediction. JAMIA Open. 2025;8(1).
- Abstract Viewed: 1 times
- 53708 Downloaded: 0 times