Machine learning models using real-world physiological time-series data can detect intraoperative anaphylaxis among patients with hypotension as early as 2-3 minutes after hypotension onset, with CatBoost achieving the best performance.
Key Findings
Results
CatBoost models achieved the best performance for early detection of intraoperative anaphylaxis on Dataset +3Min (3 minutes after hypotension onset).
On Dataset +3Min, CatBoost achieved an AUROC of 0.851 (±0.082) and AUPRC of 0.431 (±0.094)
Sensitivity was 0.861 (±0.210) and specificity was 0.783 (±0.164) on Dataset +3Min
Models were evaluated using five-fold cross-validation
Three model types were compared: random forests, extreme gradient boosting, and categorical boosting (CatBoost)
Results
CatBoost models also performed best on Dataset +2Min (2 minutes after hypotension onset), with higher specificity but lower sensitivity than the +3Min dataset.
On Dataset +2Min, CatBoost achieved an AUROC of 0.823 (±0.145) and AUPRC of 0.365 (±0.097)
Sensitivity was 0.711 (±0.183) and specificity was 0.940 (±0.065) on Dataset +2Min
The higher specificity on Dataset +2Min (0.940) compared to +3Min (0.783) suggests a trade-off between earliness of detection and discriminative balance
Results indicate the model could alert clinicians to suspected anaphylaxis just 2 minutes after hypotension
Methods
The study datasets were highly imbalanced, with anaphylactic (positive) cases comprising approximately 1 in 20 hypotensive surgical patients.
Dataset +1Min contained 49 positive samples and 980 negative samples
Dataset +2Min contained 48 positive samples and 960 negative samples
Dataset +3Min contained 44 positive samples and 880 negative samples
The positive-to-negative ratio was approximately 1:20 across all datasets
Patients were drawn from surgeries at Peking University People's Hospital between January 1, 2011 and January 1, 2023
Methods
Physiological time-series data were constructed spanning 10 minutes before to 1, 2, or 3 minutes after hypotension onset at a 10-second sampling frequency.
Data were extracted from electronic medical records (EMRs) at a 10-second sampling frequency
Three datasets were constructed: Dataset +1Min, +2Min, and +3Min, each beginning 10 minutes before hypotension
Positive groups consisted of intraoperative anaphylactic patients with hypotension; negative groups consisted of intraoperative non-anaphylactic patients with hypotension
The study period covered over 12 years (January 1, 2011 to January 1, 2023)
Conclusions
The study demonstrates the feasibility of machine learning-based early detection of intraoperative anaphylaxis using real-world physiological time-series data.
The authors state the models 'could alert clinicians to suspected anaphylaxis just 2-3 min after hypotension, enabling early detection of intraoperative anaphylaxis'
Intraoperative anaphylaxis is described as 'a rare yet fatal condition' that presents challenges due to its unpredictability
The approach leverages 'unique characteristics of physiological parameter changes that precede or occur during the early stages of anaphylaxis'
Real-world EMR data were used rather than controlled experimental data
What This Means
This research suggests that artificial intelligence can help doctors identify life-threatening allergic reactions (anaphylaxis) during surgery much sooner than currently possible. The researchers used over 12 years of patient monitoring data from a major Chinese hospital, training computer models to recognize patterns in vital signs — such as heart rate and blood pressure — that distinguish a dangerous anaphylactic reaction from other causes of low blood pressure during surgery. They tested three types of machine learning algorithms, finding that a method called CatBoost performed best, correctly identifying anaphylaxis roughly 86% of the time when given data up to 3 minutes after blood pressure dropped, while also correctly ruling it out in about 78% of non-anaphylactic cases.
The practical implication is significant: anaphylaxis during surgery is rare but can be fatal if not treated quickly, and it is notoriously difficult to distinguish from other causes of sudden low blood pressure in an already-sedated patient. This research suggests that an automated alert system using continuous monitoring data could flag suspected anaphylaxis within just 2 to 3 minutes of a blood pressure drop, giving clinicians a critical head start in administering appropriate treatment such as epinephrine.
However, there are important limitations to keep in mind. The number of confirmed anaphylaxis cases was small (44–49 patients), the data came from a single hospital in China, and the class imbalance between anaphylactic and non-anaphylactic patients was large (roughly 1 in 20), which makes the precision of the models relatively modest. Further validation in larger, more diverse patient populations would be needed before such a system could be used clinically.
Su H, You L, Jiang B, Wu Z, Kong G, Feng Y. (2026). Machine Learning-based Early Detection of Intraoperative Anaphylaxis Among Patients with Hypotension Using Real-World Physiological Time Series Data.. Journal of medical systems. https://doi.org/10.1007/s10916-026-02452-8