Image recognition of radial arterial waveforms predicts lactate levels for postoperative children in pediatric intensive care unit

Objective To evaluate a digital health framework combining image-based arterial waveform analysis and machine learning for non-invasive prediction of blood lactate levels in postoperative paediatric CHD patients. Method Seventy-seven children (39 females, 38 males; median age 3.1 months, and median weight 5.24kg) undergoing cardiac surgery were prospectively studied. During PICU admission, continuous radial arterial pressure waveforms and intermittent arterial blood samples, including lactate levels, were collected. In total, 934 paired lactate measurements and corresponding waveform segments were analysed. At each sampling time point, five-second arterial waveform images were analysed using image recognition (DeepTune software) to generate multidimensional feature vectors alongside lactate measurements. These digital biomarkers were combined with routinely available clinical and waveform-derived variables (age, body weight, height, heart rate, systolic and diastolic blood pressure, waveform sharp angle, and area under the curve) and incorporated into the machine learning pipeline. A training and fine-tuning strategy were applied to optimise the deep learning model. The final process employed a multi-stage learning framework, consisting of deep learning–based feature extraction followed by an automated machine learning–selected tabular learner. After knowledge representation embedding extraction from ResNet50, predictive performance was compared across Elastic Net, Stochastic Gradient Descent (SGD), Light Gradient Boosting Machine (LightGBM), and k-Nearest Neighbours (kNN). Model evaluation included (1) lactate classification (low <18 mg/dL, intermediate 18–36 mg/dL, high ≥36 mg/dL) assessed by accuracy, AUROC, and F1-score, and (2) regression prediction of absolute lactate values assessed by mean absolute error (MAE) and coefficient of determination (R²). Cross-validation was applied. Results Elastic Net achieved the best overall performance. For lactate classification, accuracy was 0.902, AUROC 0.963, and F1-score 0.887. For regression prediction, Elastic Net yielded R² = 0.849 and MAE = 4.255 mg/dL. Image-based arterial waveform analysis combined with a multi-stage machine learning approach enabled accurate prediction of lactate levels. Summary This digital health strategy demonstrates that AI-enabled analysis of routinely monitored arterial waveforms can accurately predict blood lactate levels in postoperative paediatric CHD patients. Such an approach may facilitate continuous, non-invasive haemodynamic surveillance, support early clinical decision-making, and reduce the need for frequent arterial blood sampling.