Optimising antibiotic switching via forecasting of patient physiology

This paper proposes a neural process-based decision support system that forecasts patient vital sign trajectories to probabilistically predict readiness for switching from intravenous to oral antibiotics, thereby outperforming random selection and historical decision-learning approaches in identifying eligible patients across US and UK datasets.

The Gist

Imagine you are a busy hospital nurse or doctor. You have a ward full of patients, many of whom are hooked up to IV drips for antibiotics. You know that switching these patients from IV drips to simple oral pills is a good thing: it makes them more comfortable, reduces infection risks, and frees up your time.

But here's the problem: You are too busy to check everyone.

In England, about 1 in 5 patients who could safely switch to pills are still stuck on IVs because no one noticed they were getting better. The doctors are overwhelmed, and the "switch" decision gets missed.

This paper proposes a smart AI assistant to help solve this. Instead of trying to guess what doctors did in the past (which might have been slow or inconsistent), this AI acts like a weather forecaster for the human body.

Here is how it works, broken down into simple concepts:

1. The "Weather Forecast" Analogy

Imagine you want to know if it's safe to go for a picnic tomorrow.

• The Old Way (Learning from History): You look at what people did last year. "Oh, last year we stayed inside on Tuesdays, so we should stay inside today." This is flawed because last year might have been an anomaly, or the weather might have changed.
• The New Way (Forecasting): You look at the clouds, the wind, and the pressure right now and predict what the weather will be like tomorrow. If the forecast says "Sunny and calm," you know it's a good day for a picnic.

The AI does exactly this for patients. It looks at a patient's current vital signs (heart rate, temperature, breathing, etc.) and uses a special math model called a Neural Process to predict what those signs will look like over the next 12 hours.

2. The "Traffic Light" System

Once the AI predicts the future, it doesn't just give a number; it checks the "rules of the road."

• Green Light: The forecast says the patient's heart rate, temperature, and breathing will stay within safe, normal ranges. Result: "This patient is ready to switch to pills!"
• Red Light: The forecast predicts a spike in fever or a drop in blood pressure. Result: "Wait, they aren't stable yet."

The AI then creates a ranked list for the doctors. It says, "Here are the top 10 patients who are most likely to be ready for a switch today." This helps the busy medical team focus their attention where it matters most.

3. Why is this better than other AI?

Most AI systems learn by watching what humans do. If doctors were slow to switch patients in the past, the AI learns to be slow too. It's like a student who only copies the teacher's mistakes.

This system is different. It doesn't care what the doctor did yesterday. It only cares about the physics of the patient's body.

• It asks: "Is the body getting better?"
• It ignores: "Did the doctor switch them last time?"

This means if medical guidelines change tomorrow (e.g., "We now know patients can switch earlier"), you don't need to retrain the AI. You just change the "traffic light" rules, and the AI instantly adapts.

4. The "Crystal Ball" with a Safety Net

The AI is probabilistic, which is a fancy way of saying it gives a confidence score.

• Instead of saying "Yes, switch," it says, "There is an 88% chance this patient will be stable in 12 hours."
• This is crucial because doctors need to trust the tool. They can look at the forecast and say, "Okay, the AI thinks the heart rate will stay low, but I see a small spike in the prediction. I'll double-check before switching."

5. The Results

The team tested this on two huge groups of patients: one in the US (very sick ICU patients) and one in London (a mix of general hospital patients).

• The Magic: The AI found patients ready to switch 2 to 3 times better than if a nurse just picked patients at random.
• The Goal: It doesn't replace the doctor. It's a "spotter" that highlights the patients who are ready, so the doctor can make the final call with confidence.

Summary

Think of this system as a smart, tireless assistant that watches the vital signs of every patient, predicts their "health weather" for the next day, and hands the doctor a list of patients who are sunny and ready to go home on pills. It saves time, reduces infections, and ensures no patient is left on an IV drip longer than necessary.

Technical Summary

Here is a detailed technical summary of the paper "Optimising antibiotic switching via forecasting of patient physiology."

1. Problem Statement

The timely transition of hospital patients from intravenous (IV) to oral antibiotic therapy (IVOS) is critical for reducing hospital stays, preventing catheter-related infections, and lowering costs. However, in England, approximately 19% of patients remain on IV antibiotics despite meeting clinical switching criteria.

Current Clinical Decision Support Systems (CDSS) face two major limitations:

1. Rule-based systems: Rely on fixed thresholds that fail to capture the trend of patient improvement (e.g., a patient might meet a threshold at one moment but be unstable overall).
2. Supervised learning models: Trained on historical clinical decisions (EHR labels). These models inevitably learn and reproduce existing suboptimal practices, delays, and inconsistencies rather than improving upon clinical standards. Furthermore, constructing reliable "ground truth" labels for switching is difficult due to complex, non-linear antibiotic prescribing patterns and the multifaceted nature of the decision (which involves factors beyond vital signs).

The authors propose a system that does not learn from past actions but instead forecasts future physiology to determine switch-readiness.

2. Methodology

A. Core Approach: Forecasting Physiology

Instead of predicting a binary "switch" label directly, the system forecasts the trajectory of key vital signs over a future window (e.g., the next 12 hours).

• Input: Irregularly sampled historical vital signs (Heart Rate, Respiratory Rate, SpO2, Systolic Blood Pressure, Temperature).
• Mechanism: A Neural Process (NP) model, specifically a Convolutional Conditional Neural Process (ConvCNP), is used. This probabilistic model handles irregular sampling and missing data inherent in Electronic Health Records (EHRs).
• Output: A probabilistic distribution of future vital sign values, rather than a single point estimate.

B. Switch-Readiness Determination

The system applies clinical criteria after forecasting:

1. Forecasting: The model predicts vital sign trajectories for a future horizon (e.g., 12 hours).
2. Criteria Application: The predicted distributions are checked against predefined clinical normal ranges.
3. Probability Calculation: The probability that all vital signs will remain within normal ranges throughout the forecast window is calculated. This probability serves as the "switch-readiness" score.
4. Ranking: Patients are ranked by this probability, allowing clinicians to prioritize reviews for those most likely to be ready for switching.

C. Model Architecture & Baselines

• Proposed Model: ConvCNP with a UNet architecture. It outputs Gaussian predictive distributions. A "Tuned" variant (NP-tuned) adds a classification head to improve probability calibration.
• Baselines:
  • Repeat: Uses the last observed value as the prediction.
  • GBDT (Gradient Boosted Decision Trees): Uses statistical features (mean, trend, frequency) from the look-back window.
  • Logistic Regression: Direct classification baseline.
• Datasets: Validated on MIMIC-IV (US ICU, 6,333 encounters) and UCLH (UK General Hospital, 10,584 encounters). These represent distinct populations and clinical practices.

3. Key Contributions

1. Paradigm Shift from Action to Physiology: The paper proposes forecasting patient physiology rather than learning historical switching decisions. This avoids the "learning suboptimal behavior" trap and allows the system to adapt to updated clinical guidelines without retraining the forecasting model.
2. Probabilistic Forecasting for Decision Support: By using Neural Processes, the system provides a full predictive distribution. This allows for the calculation of the probability that a patient will remain stable, offering a more nuanced and interpretable metric than point estimates.
3. Interpretability and Flexibility:
   • Clinicians can inspect the underlying vital sign forecasts to understand why a patient is flagged.
   • The system can easily adapt to specific clinical contexts (e.g., ignoring respiratory rate for ventilated patients) by adjusting criteria post-forecasting, without retraining the core model.
4. Robust Validation: The approach was validated across two distinct datasets (US ICU vs. UK General Hospital), demonstrating generalizability across different patient acuity levels and healthcare systems.

4. Results

A. Forecasting Performance

• The Neural Process (NP) model significantly outperformed the Repeat and GBDT baselines in Mean Absolute Error (MAE) for most vital signs across both datasets.
• On the more complex MIMIC-IV dataset, NP improved MAE by ~3% over GBDT.
• Performance degraded slightly as the forecast horizon increased (6h vs 12h vs 24h), but NP remained superior or comparable to baselines.

B. Switch-Readiness Prediction (Ranking)

The primary goal is to rank patients effectively for review.

• Precision@5: This metric measures the proportion of the top 5 ranked patients who are truly ready to switch.
  • UCLH: NP selected 2.2× more relevant patients than random selection.
  • MIMIC: NP selected 3.2× more relevant patients than random selection.
• Comparison: The NP model and GBDT performed similarly in ranking (AUROC and Average Precision), with NP showing slight advantages in calibration and handling uncertainty.
• Calibration: The standard NP model showed some calibration issues (underestimating probabilities), which were resolved in the "NP-tuned" variant.

C. Case Study

A visual case study (Figure 5) demonstrated that the model correctly tracked a patient's physiological improvement over two weeks, assigning a rising probability of switch-readiness as vital signs stabilized, even though the patient was not switched in reality (suggesting a missed opportunity for earlier switching).

5. Significance and Limitations

Significance:

• Clinical Utility: The system acts as a "triage" tool, surfacing the most promising candidates for expert review rather than automating the decision, thereby preserving clinical judgment.
• Adaptability: Because the model forecasts physiology and criteria are applied separately, the system can adapt to new evidence or guidelines (e.g., changing threshold values) without the need for expensive and time-consuming model retraining.
• Addressing Data Bias: By not relying on historical switching labels (which are noisy and biased), the system avoids reinforcing current inefficiencies in antibiotic stewardship.

Limitations:

• Retrospective Evaluation: The study lacks prospective clinical trials. While the model ranks patients well, its actual impact on reducing IV days or improving patient outcomes has not yet been measured in a live setting.
• Scope of Variables: The model relies solely on five vital signs. It does not incorporate inflammatory markers (e.g., CRP, WBC) or other clinical factors (e.g., source control, wound status) which are crucial for the final decision.
• Ground Truth Proxy: The "switch-ready" label is a proxy based on vital signs. It does not account for oral suitability (e.g., patient ability to swallow, drug availability), which remains the clinician's responsibility.

Conclusion

The paper demonstrates that forecasting patient physiology using probabilistic Neural Processes provides a principled, interpretable, and adaptable foundation for antibiotic stewardship. By shifting the focus from learning historical actions to predicting future physiological stability, the system offers a robust method to prioritize patients for IV-to-oral switching, potentially reducing unnecessary IV therapy days and improving hospital efficiency.