Optimising antibiotic switching via forecasting of patient physiology

Magnus Ross; Nel Swanepoel; Akish Luintel; Emma McGuire; Ingemar J. Cox; Steve Harris; Vasileios Lampos

cs.LGMar 9, 2026

Optimising antibiotic switching via forecasting of patient physiology

Authors: Magnus Ross, Nel Swanepoel, Akish Luintel, Emma McGuire, Ingemar J. Cox, Steve Harris, Vasileios Lampos

Organizations: Department of Computer Science, Centre for AI, UCL, UK · Institute of Health Informatics, UCL, UK · Department of Medical Microbiology, Infection Division, University College Hospital London, UK · Department of Computer Science, University of Copenhagen, Denmark

Abstract

Timely transition from intravenous (IV) to oral antibiotic therapy shortens hospital stays, reduces catheter-related infections, and lowers healthcare costs, yet one in five patients in England remain on IV antibiotics despite meeting switching criteria. Clinical decision support systems can improve switching rates, but approaches that learn from historical decisions reproduce the delays and inconsistencies of routine practice. We propose using neural processes to model vital sign trajectories probabilistically, predicting switch-readiness by comparing forecasts against clinical guidelines rather than learning from past actions, and ranking patients to prioritise clinical review. The design yields interpretable outputs, adapts to updated guidelines without retraining, and preserves clinical judgement. Validated on MIMIC-IV (US intensive care, 6,333 encounters) and UCLH (a large urban academic UK hospital group, 10,584 encounters), the system selects 2.2-3.2 $\times$ more relevant patients than random. Our results demonstrate that forecasting patient physiology offers a principled foundation for decision support in antibiotic stewardship.

Figures & tables

Figure 1: The proposed AI-enabled IVOS system. A) Training a probabilistic neural process model to forecast patient trajectories. B) Using the trained model to generate likely trajectories over the following day and applying clinical criteria to determine the proportion of samples ready to be switched. C) Presentation of switch-readiness probabilities as a ranked list of patients for review.

Statistic	MIMIC	UCLH
Number of admissions	6,333	10,584
Unique subjects	5,850	8,701
Age (years)	65.70 $\pm$ 15.29	55.98 $\pm$ 19.24
Length of stay (days)	17.69 $\pm$ 17.07	14.23 $\pm$ 14.42
Sex
Female	2,579 (40.7%)	5,564 (52.6%)

Table 1: Comparison of the evaluation set cohorts for MIMIC-IV and UCLH datasets. Categorical variables are reported as count (percentage of admissions). Continuous variables are reported as mean

\pm

standard deviation. MIMIC-IV records race while UCLH records ethnicity, reflecting the coding systems used in each source database; “Other White Bg” denotes Other White Background. Antibiotic frequencies may sum to more than 100% as patients can receive multiple antibiotics during an admission. The bottom section shows the mean number of vital sign measurements recorded per admission for each variable. Abbreviations: Pip.-Taz. (Piperacillin-Tazobactam), Sulfameth/Trimeth (Sulfamethoxazole-Trimethoprim), IV (Intravenous), SpO 2 (Oxygen Saturation), BP (Blood Pressure).

Dataset	Model	Variable
		HR	RR	SpO 2	SBP	Temp.
		(bpm)	(bpm)	(%)	(mmHg)	( ∘ F)
UCLH	Repeat	8.79 [8.76-8.82]	1.97 [1.96-1.98]	1.51 [1.51-1.52]	12.86 [12.81-12.91]	0.69 [0.69-0.70]
	GBDT	7.94 [7.92-7.96]	1.76 [1.76-1.77]	1.31 [1.30-1.31]	10.86 [10.83-10.89]	0.58 [0.58-0.58]
	NP	7.58 [7.56-7.60]	1.71 [1.70-1.72]	1.30 [1.30-1.31]	10.71 [10.67-10.75]	0.57 [0.57-0.57]
MIMIC	Repeat	9.17 [9.14-9.19]	4.09 [4.08-4.11]	2.05 [2.04-2.05]	14.29 [14.24-14.34]	0.63 [0.62-0.63]

Table 2: Mean absolute error of model forecasts for 12 hour prediction horizon. The 95% confidence intervals are shown, computed by a bootstrap resampling of the mean error over each of the windows in the test set. Bold values indicate the best performing model(s) (p

\geq

0.05). Variable abbreviations: HR (Heart Rate), RR (Respiratory Rate),

\text{SpO}_{2}

(Oxygen Saturation), SBP (Systolic Blood Pressure), Temp. (Temperature). Model abbreviations: GBDT (Gradient Boosted Decision Trees), NP (Neural Process).

Figure 2: Forecasting performance on the MIMIC data for each variable with increasing forecast horizon for NP model and baselines. Note horizontal axis scale is logarithmic. Error bars show 95% confidence. Variable Abbreviations: HR (Heart Rate), RR (Respiratory Rate),

\text{SpO}_{2}

(Oxygen Saturation), SBP (Systolic Blood Pressure), Temp. (Temperature). Model abbreviations: GBDT (Gradient Boosted Decision Trees), NP (Neural Process).

Dataset	Model	Metric
		AUROC	AP	Brier Score	Precision@5
UCLH	Logistic	0.827 [0.823-0.831]	0.740 [0.732-0.747]	0.167 [0.165-0.169]	0.836 [0.823-0.850]
	GBDT	0.862 [0.858-0.865]	0.794 [0.787-0.800]	0.149 [0.147-0.151]	0.899 [0.888-0.910]
	NP	0.858 [0.854-0.861]	0.785 [0.779-0.792]	0.216 [0.212-0.219]	0.883 [0.871-0.894]
	NP-tuned	0.867 [0.864-0.871]	0.800 [0.793-0.807]	0.146 [0.144-0.148]	0.901 [0.890-0.912]
MIMIC	Logistic	0.759 [0.749-0.767]	0.300 [0.284-0.315]	0.098 [0.095-0.100]	0.280 [0.263-0.297]

Table 3: Comparison of model performance for switch-readiness prediction. Higher values are better for all metrics except the Brier score. Bold values indicate the best performing model(s) (p

\geq

0.05). Metric abbreviations: AUROC (Area under the Receiver Operating Characteristic), AP (Average Precision). Model abbreviations: GBDT (Gradient Boosted Decision Trees), NP (Neural Process).

Panel of Figure 3 — Figure 3: Comparison of model performance curves for both datasets. From left to right the plots show Receiver Operating Characteristic, precision-recall, and calibration curves. For the former two the dashed line shows the performance of a chance model, for the latter it shows a perfect model. For the calibration curve, aggregation bins are the quantiles of the distribution of predicted probabilities; the small proportion of positive examples in MIMIC means the curves do not span the full range of the plot. The shaded region indicates 95% confidence. Model abbreviations: GBDT (Gradient Boosted Decision Trees), NP (Neural Process).

Figure 4: Ranking performance on the MIMIC data for each variable with increasing forecast horizon for NP models and baselines. Horizontal axis scale is logarithmic. Error bars show 95% confidence. Note that the classes become more imbalanced as the forecast horizon increases, with 20.1%, 12.2%, and 7.8% of the instances being positive for the 6, 12, and 24 hour horizons respectively. Model abbreviations: GBDT (Gradient Boosted Decision Trees), NP (Neural Process).

Dataset	Model	Metric
		F1 Score	Accuracy	Precision	Recall
UCLH	Repeat	0.704 [0.699-0.709]	0.700 [0.696-0.704]	0.587 [0.581-0.594]	0.879 [0.874-0.884]
	Logistic	0.719 [0.714-0.724]	0.736 [0.732-0.740]	0.634 [0.627-0.640]	0.832 [0.827-0.838]
	GBDT	0.744 [0.740-0.749]	0.765 [0.761-0.769]	0.667 [0.661-0.673]	0.843 [0.837-0.848]
	NP	0.741 [0.736-0.746]	0.754 [0.751-0.759]	0.649 [0.643-0.655]	0.863 [0.857-0.868]
	NP-tuned	0.748 [0.744-0.754]	0.764 [0.760-0.768]	0.660 [0.654-0.666]	0.864 [0.859-0.870]

Table 4: Comparison of model performance for binary classification, with the positive class representing a patient that is ready to be switched. The proportion of positive instances is 40.2% and 12.2% for UCLH and MIMIC respectively, indicating predicting the majority class could achieve an accuracy of 59.8% and 87.8% in each case. Decision thresholds were chosen to maximise the F1 score on the validation set. 95% confidence intervals are shown in brackets. Bold values indicate the best performing model(s) (p

\geq

0.05). A higher value is better for all metrics. Model abbreviations: GBDT (Gradient Boosted Decision Trees), NP (Neural Process).

Figure 5: Example encounter from the MIMIC test set illustrating the intended functioning of the system. The panels show (from top to bottom): predicted switch-readiness probabilities from the NP-tuned model; labels indicating oral (PO) and IV antibiotic prescriptions and whether clinical criteria are satisfied in each forecast window; measured vital signs with clinical thresholds for switch-readiness overlaid; and antibiotic prescription timelines, with colour and pattern indicating administration route. Grey vertical lines demarcate daily 12-hour forecast windows. Forecasts begin 48 hours after admission to ensure sufficient historical data for prediction. Variable Abbreviations: HR (Heart Rate), RR (Respiratory Rate),

\text{SpO}_{2}

(Oxygen Saturation), SBP (Systolic Blood Pressure), Temp. (Temperature). Antibiotic Abbreviations: Sulfameth/Trimeth (Sulfamethoxazole-Trimethoprim).

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Figure S1 : Left plot shows calibration curve for the forecasting model’s predicted uncertainty. For each confidence level (e.g., 80%), we generate a prediction interval that should contain the true value with that probability using the Normal predictive distribution. The x-axis shows the chosen confidence level while the y-axis shows the actual proportion of test observations that fell within the intervals. Perfect calibration follows the diagonal: an 80% interval should contain 80% of observations. Points above the diagonal indicate underconfident predictions. Right shows relationship between predicted uncertainty and absolute prediction error. Individual predictions are shown as scatter points, with binned averages over 1st to 99th percentiles (red line) summarising the trend. The diagonal represents ideal calibration where predicted uncertainty magnitude matches actual error magnitude. In this case, an underconfident model has binned average error below the diagonal; where the predicted uncertainty is higher than the actual error.

Variable	Range	Units
Heart Rate	10 – 400	bpm
Systolic Blood Pressure	0 – 400	mmHg
Oxygen Saturation	0 – 100	%
Respiration Rate	0 – 120	bpm
Temperature	50 – 120	°F

Appendix

Table S1 : Ranges of plausible values used for data cleaning.

	MIMIC-IV		UCLH
Partition	Forecasting	Switch-readiness	Forecasting	Switch-readiness
Train	167,923	79,148	344,133	76,586
Val.	18,804	8,902	38,864	8,885
Test	52,537	22,615	186,759	41,053

Appendix

Table S2 : Number of tasks in each dataset partition for the forecasting and switch-readiness prediction tasks.

Measurement	Switch Criterion
Temperature ( ∘ F)	96.8 – 100.4
Heart Rate (bpm)	41 – 90
Respiratory Rate (breaths/min)	9 – 20
Systolic BP (mmHg)	101 – 219
SpO2 (%)	$>94$

Appendix

Table S3 : Physiological criteria for switch readiness.

Architecture
Parameter	Value
Number of layers	4
Number of channels ∗	{64, 128, 256, 512}
Lengthscale	1.0 hour
Training
Learning rate ∗	{ $10^{-5}$ , $5\times 10^{-5}$ , $10^{-4}$ , $5\times 10^{-4}$ , $10^{-3}$ }

Appendix

Table S4 : Neural process model hyperparameters. Parameters marked with ∗ were selected via grid search, with the search space shown in parentheses.

Architecture
Parameter	Value
Hidden layer 1 size	256
Hidden layer 2 size	128
Dropout rate	{0.0, 0.1, 0.5}
Training
Learning rate	{ $5\times 10^{-5}$ , $10^{-4}$ , $5\times 10^{-4}$ , $10^{-3}$ }

Appendix

Table S5 : NP-tuned classification head hyperparameters. The classification head is trained with the base forecasting model weights frozen.

Parameter	Value
GBDT
Number of estimators	{5, 10, 50, 100, 200}
Maximum depth	{3, 4, 5, 6}
Learning rate	{0.001, 0.005, 0.05, 0.1, 0.5}
Logistic Regression
Inverse regularization strength (C)	{0.1, 1.0, 10.0, 100.0, 1000.0}

Appendix

Table S6 : Baseline model hyperparameters. A grid search was performed over the sets of hyperparameters. The model with the highest average precision on the validation set was selected for evaluation on the test set.

Measurement	Switch Criterion
Temperature ( ∘ F)	96.8 – 100.58
Heart Rate (bpm)	40 – 131
Respiratory Rate (breaths/min)	8 – 24
Systolic BP (mmHg)	90 – 229
SpO2 (%)	$>91$

Appendix

Table S7 : Alternative, loose criteria for switch readiness.

Dataset	Model	Metric
		AUROC	AP	Brier Score	Precision@5
MIMIC-loose	Logistic	0.676 [0.669-0.682]	0.650 [0.640-0.660]	0.229 [0.228-0.231]	0.680 [0.663-0.696]
	GBDT	0.829 [0.824-0.834]	0.809 [0.801-0.817]	0.168 [0.165-0.170]	0.835 [0.821-0.849]
	NP	0.836 [0.831-0.841]	0.810 [0.802-0.817]	0.239 [0.235-0.243]	0.841 [0.827-0.855]
	NP-tuned	0.814 [0.809-0.820]	0.792 [0.783-0.800]	0.175 [0.172-0.177]	0.822 [0.808-0.838]

Appendix

Table S8 : Model performance for switch-readiness prediction on the MIMIC dataset with looser clinical criteria (switch rate 50.0%). Bold values indicate the best performing model(s) (p

\geq

0.05). Model abbreviations: GBDT (Gradient Boosted Decision Trees), NP (Neural Process).

Dataset	Model	Metric
		F1 Score	Accuracy	Precision	Recall
MIMIC-loose	Repeat	0.739 [0.733-0.745]	0.700 [0.695-0.706]	0.654 [0.647-0.662]	0.849 [0.842-0.856]
	Logistic	0.684 [0.678-0.690]	0.560 [0.554-0.567]	0.534 [0.527-0.541]	0.950 [0.946-0.954]
	GBDT	0.772 [0.766-0.777]	0.741 [0.736-0.747]	0.690 [0.683-0.697]	0.875 [0.869-0.881]
	NP	0.779 [0.773-0.784]	0.753 [0.748-0.759]	0.707 [0.699-0.714]	0.867 [0.861-0.872]
	NP-tuned	0.762 [0.756-0.768]	0.730 [0.724-0.736]	0.681 [0.674-0.689]	0.865 [0.859-0.872]

Appendix

Table S9 : Model performance for binary switch-readiness prediction on the MIMIC dataset with looser clinical criteria (switch rate 50.0%). Bold values indicate the best performing model(s) (p

\geq

0.05). Model abbreviations: GBDT (Gradient Boosted Decision Trees), NP (Neural Process).

Figure S2 : Example encounter from the MIMIC test set illustrating a case in which the system’s prediction is less clearly beneficial. The system recommends an earlier switch, but then the patient becomes more unwell as they are switched. In this encounter the test days only span around half of the full stay, since the IV prescription is halted after 150 hours. Panels show the same information as those in the figure in the main text. Variable Abbreviations: HR (Heart Rate), RR (Respiratory Rate),

\text{SpO}_{2}

(Oxygen Saturation), SBP (Systolic Blood Pressure), Temp. (Temperature). Antibiotic Abbreviations: Pip.-Taz., Piperacillin-Tazobactam; Sulfameth/Trimeth, Sulfamethoxazole-Trimethoprim

Figure S3 : Example encounter from the MIMIC test set. In this example the patient is unwell at the very start of the encounter, but their vitals are stable for most of the remainder, aside from a few spikes. The model becomes more confident the patient is ready to be switched as the encounter progresses, although no switch occurs. Panels show the same information as those in the figure in the main text. Variable Abbreviations: HR (Heart Rate), RR (Respiratory Rate),

\text{SpO}_{2}

(Oxygen Saturation), SBP (Systolic Blood Pressure), Temp. (Temperature).

Clinical Decision Support

Clinical Decision Support uses data, medical knowledge, and AI to help clinicians assess patients and choose tests, diagnoses, and treatments.

See all 242 papers

Benchmarking Machine Learning Architectures for Antimicrobial Stewardship in Pediatric ICUs
May 21, 2026Niklas Raehse, Luregn J. Schlapbach, Daphné ChopardClinical Prediction Longitudinal EHR Modeling
Department of Intensive Care and Neonatology and Children’s Research Center, University of Zurich, University Children’s Hospital Zurich, Zurich, Switzerland · Department of Computer Science, ETH Zurich, Zurich, Switzerland
Audit Before You Update: Recording Artifacts and the Limits of Bounded Model Updates in MIMIC-IV ICU Prediction
Jul 21, 2026Fatema Ferdous Tamanna, K. M. Merajul Arefin, Md. Abdul MasudClinical Decision Support Longitudinal EHR Modeling
Dept. of Computer Science and IT, Patuakhali Science and Technology University, Bangladesh · Dept. of Computer Science and Engineering, University of Dhaka, Bangladesh
DRIFT: Direct-Recursive Intervention-Conditioned Forecasting of ICU Physiological Trajectories
Jul 28, 2026Weixin Liu, Juming Xiong, Congning Ni +4Multivariate Time Series Forecasting Time Series Forecasting
Vanderbilt University, Nashville, Tennessee, USA · Vanderbilt University Medical Center, Nashville, Tennessee, USA

May 21, 2026J/K move · Enter open · S save

Benchmarking Machine Learning Architectures for Antimicrobial Stewardship in Pediatric ICUs

Niklas Raehse, Luregn J. Schlapbach, Daphné Chopard

Department of Intensive Care and Neonatology and Children’s Research Center, University of Zurich, University Children’s Hospital Zurich, Zurich, Switzerland · Department of Computer Science, ETH Zurich, Zurich, Switzerland

Clinical Prediction Longitudinal EHR Modeling

Open paper AskPDF arXiv

Antimicrobial stewardship (AMS) is critical in pediatric intensive care units (PICUs), where diagnostic uncertainty often drives broad-spectrum antibiotic use, increasing antimicrobial resistance and potential long-term harms. Machine learning offers a promising approach for identifying patient-level opportunities for stewardship interventions from electronic health record data, yet prior work has focused largely on adult populations and static tabular representations. We present a systematic benchmarking study of AMS intervention prediction in the PICU across the public Paediatric Intensive Care database a private cohort from the University Children's Hospital Zurich, Switzerland. We define four clinically relevant proxy targets for reducing antibiotic exposure: intravenous-to-oral switching, de-escalation, discontinuation, and short-course therapy. Under a unified evaluation framework, we compare tabular, sequence-based, and graph-based temporal models at multiple temporal resolutions. We find that predictive performance is driven primarily by target prevalence and dataset characteristics rather than model complexity. Sequence models improve the precision-recall trade-off over tabular approaches at coarse (24-hour) resolution, while finer temporal modeling provides limited additional benefit. However, these gains come at the cost of poorer calibration, with simpler tabular models yielding more reliable probability estimates. Our findings highlight the importance of target design, temporal representation, and calibration in clinical machine learning, and provide practical guidance for developing reliable decision support systems for pediatric AMS.

Abstract

Clinical Decision Support

Explore similar work

Benchmarking Machine Learning Architectures for Antimicrobial Stewardship in Pediatric ICUs

Audit Before You Update: Recording Artifacts and the Limits of Bounded Model Updates in MIMIC-IV ICU Prediction

DRIFT: Direct-Recursive Intervention-Conditioned Forecasting of ICU Physiological Trajectories