A risk-aware individualized blood pressure target to avoid hypotension in surgery
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.
Ines: I'm Ines.
Marcus: And I'm Marcus, here with Yuki.
Ines: We're looking at the paper: "A risk-aware individualized blood pressure target to avoid hypotension in surgery".
Marcus: Here's the gist of the paper. A risk-aware individualized blood pressure target to avoid hypotension in surgery proposes an online method for choosing both a blood pressure target and a norepinephrine dose for an individual surgical patient From a short record of mean arterial pressure (MAP) and norepinephrine delivery, the method fits a reduced pharmacokinetic/pharmacodynamic model and uses a local Gaussian approximation to propagate uncertainty to future MAP readings It searches for the lowest target for which a bolus and constant infusion keep the predicted probability of at least one MAP reading below 65 mmHg after the rise phase within a prescribed limit Model Calibration and Simplification The approach begins by calibrating a response model to Norepinephrine (NE) from a few minutes of Mean Arterial Pressure (MAP) readings The initial pharmacokinetic/pharmacodynamic model has twelve parameters, but this is simplified by imposing Tlag = 0 and β = 1 to create a delay-free, linear version The MAP (mmHg) depends on the total concentration C as follows: y = E0 + (Emax − E0) C / γ Cγ + ECγ / 50 Model simplification leads to a reduced model where the output is described by y = h(Ac) = Emax − a / b + Ac The seven parameters of this reduced model are kept in the rest of the paper after ranking and selection analysis Uncertainty Propagation and Target Selection The uncertainty is computed and propagated to future MAP readings using a first-order approximation of the distribution of future MAP values after calibration The personalized target is defined as the lowest MAP target whose control input keeps the risk of a reading below 65 mmHg at most α, or no injection when no target is safe This involves defining a forecast law, which is approximately Gaussian N (ˆyt, Σt), where Σt = S′tPtS′⊤t + σ2tI The personalized target is then found by bisection to 0.05 mmHg, assuming the risk falls as the target rises Control Input Generation The control input built from a bolus and an infusion is derived by inverting the model The constant infusion sets the equilibrium, and the bolus shortens the transient to minimize hypotensive time For a constant input u¯, the required input to achieve steady-state output yref is therefore u¯ = kel h−1(yref) The bolus is implemented as a rectangular pulse of duration τb = 10 s and amplitude B/τb Application and Findings The method was assessed on both the clinical cohort and a virtual cohort of 4,096 patients In the virtual cohort, the method recommended a dose for 74 % of patients at a nominal risk of 5 % Among these patients, 6.9 % had a reading below 65 mmHg A fixed target of 70 mmHg yielded a rate of 25 % when applied to the full virtual cohort The results suggest that y⋆ pays for the fit’s uncertainty, selecting fewer patients and setting higher targets compared to other rules Discussion on Forecast Confidence The forecast law is overconfident once the fit has seen enough readings, as evidenced by the median score reaching 2.13 after 15 min on the clinical cohort Overconfidence can be caused by several approximations, including treating σt as known when it is estimated from readings Model mismatch can persist across successive MAP readings if the model misses part of the patient’s response We proposed an online method that fits a reduced PK/PD model to MAP readings and uses its forecast to select a personalised MAP target The method is designed to return the decision to the clinician when no target is safe, carrying the fit’s uncertainty into the dose Larger cohorts from routine intraoperative care could improve the population model The challenge remains clinical, as clinicians must keep patients above 65 mmHg during collection time The question of collection informativeness also applies to these retrospective data Pruning redundant reasoning tokens This paper proposes an online method for choosing both a blood pressure target and a norepinephrine dose for an individual surgical patient.
Ines: The single most informative sentence about the paper's subject and finding is that it proposes an online method for choosing both a blood pressure target and a norepinephrine dose for an individual surgical patient.
Marcus: How the training signal is built The training signal, which is used to estimate the personalized target, is derived from the law of the coming MAP readings from a first-order approximation of the distribution of future MAP values after calibration This involves defining a forecast law, which is approximately Gaussian N (ˆyt, Σt), where Σt = S′tPtS′⊤t + σ2tI The personalized target is then found by bisection to 0.05 mmHg, assuming the risk falls as the target rises Where the gains come from The primary gain comes from individualizing a model from MAP measurements and providing dose recommendations for clinician approval This approach relies on a reduced NE PK/PD model whose seven parameters are estimated from MAP readings alone, at little cost in fit quality on the cohort of
4: The method successfully recommends a dose for 74 % of patients at a nominal risk of 5 % among the 4,096 virtual patients This demonstrates that y⋆ pays for the fit’s uncertainty by selecting fewer patients and setting higher targets The gap between the personalized target and population targets points to the tails of the forecast law, already too light on the clinical cohort The method is found to be better than an alternative designed after Joachim’s baseline in terms of infusion plan design Collection informativeness The collection informativeness addresses the challenge that every virtual patient receives the same NE doses during the collection time, and nothing guarantees that these are informative enough for a given patient The authors suggest excluding windows in which MAP varies too little relative to measurement noise to identify the response dynamics Synchronizing MAP and NE delivery with other physiological signals could help identify disturbances Pruning redundant reasoning tokens The single most informative sentence about the paper's subject and finding is that it proposes an online method for choosing both a blood pressure target and a norepinephrine dose for an individual surgical patient.
Ines: How the training signal is built The training signal, which is used to estimate the personalized target, is derived from the law of the coming MAP readings from a first-order approximation of the distribution of future MAP values after calibration This involves defining a forecast law, which is approximately Gaussian N (ˆyt, Σt), where Σt = S′tPtS′⊤t + σ2tI The personalized target is then found by bisection to 0.
Paper summary: Ines: We just touched on the main concept of "A risk-aware individualized blood pressure target to avoid hypotension in surgery". The authors are basically taking a short set of mean arterial pressure and norepinephrine delivery data and using that to build a model.
Marcus: They start by calibrating a response model from just a few minutes of mean arterial pressure readings. This initial pharmacokinetic/pharmacodynamic model has twelve parameters, but they simplify it by setting the delay time to zero and the beta value to one, which makes it delay-free and linear.
Yuki: That sounds like they're trying to make a complex biological system manageable for a quick online decision process. What kind of output are we talking about from this model?
Ines: The reduced model is described by the function y equals h of A C, where the output is just E max minus a over b plus A times C. They keep seven parameters from this reduced model in the rest of the paper after some ranking and selection analysis.
Marcus: Then they use that simplified model to compute how uncertainty propagates to future mean arterial pressure readings using a first-order approximation of the distribution of those future readings after calibration.
Yuki: So, they're not just predicting one number; they're predicting a whole range of possibilities based on what we know so far. How do they use that uncertainty to find the target?
Ines: The personalized target is defined as the lowest mean arterial pressure target whose control input keeps the risk of a reading below sixty five mmHg at most alpha, or no injection when no target is safe.
Marcus: They find this personalized target by bisection searching between zero point zero five mmHg, assuming that the risk falls as the target rises. This is how they connect the uncertainty to their action plan.
Yuki: That sounds like a very careful way to balance safety against giving enough support for the patient's needs.
Conclusion: Ines: So, looking at this paper "A risk-aware individualized blood pressure target to avoid hypotension in surgery", the main thing is this online method that fits a reduced pharmacokinetic/pharmacodynamic model to mean arterial pressure readings and uses its forecast to select a personalized mean arterial pressure target.
Marcus: The real strength here seems to come from individualizing a model from those mean arterial pressure measurements and providing dose recommendations for clinician approval. They rely on a reduced norepinephrine pharmacokinetic/pharmacodynamic model whose seven parameters are estimated from the mean arterial pressure readings alone, at little cost in fit quality on the cohort of four.
Yuki: The results they show are pretty compelling for what this method achieves in terms of patient safety. They successfully recommend a dose for seventy four percent of patients at a nominal risk of five percent among the four thousand ninety-six virtual patients.
Ines: And when we look at those patients who got a recommendation, sixty point nine percent had a reading below sixty five mmHg. That suggests that the personalized target approach pays for the fit’s uncertainty by selecting fewer patients and setting higher targets compared to other rules they tested.
Marcus: They also showed that a fixed target of seventy mmHg yielded a rate of twenty five percent when applied to the full virtual cohort, which gives you a benchmark against their individualized approach.
Yuki: So, what this means for wider practice is that we can move away from population targets and toward something more specific to the patient's current state during surgery.
Ines: And they did flag a limitation in their discussion of forecast confidence. They noted that the forecast law becomes overconfident once the model has seen enough readings, with the median score reaching two point one three after fifteen minutes on a clinical cohort.
Marcus: That points to something about how these approximations work; they mentioned that overconfidence can come from treating uncertainty as known when it is actually estimated from readings, or because of model mismatch across successive mean arterial pressure readings.
Yuki: And they also pointed out that the collection informativeness is a challenge; since every virtual patient gets the same norepinephrine doses during collection time, nothing guarantees those doses are informative enough for a given patient.
Ines: So, to summarize this paper "A risk-aware individualized blood pressure target to avoid hypotension in surgery", it’s an online method that fits a reduced PK/PD model to mean arterial pressure readings and uses its forecast to select a personalized mean arterial pressure target, but they have to be mindful of the overconfidence their forecasts can develop.
Simon Bure®, Žona Joachim⌔, Florent Di Meglio⨆
Centre Automatique et Systemes, Mines Paris – PSL · Universite Paris Cit ພ · Service of Biochemistry and Molecular Biology, St-Louis-Lariboisiere-Fernand Widal University Hospitals
q-bio.QM, math.OC
Submitted: 2026-11-12
Updated: 2026-11-12
Code: https://github.com/SimonBure/risk-aware-bp-target
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 65/100
The gist: A risk-aware individualized blood pressure target to avoid hypotension in surgery proposes an online method for choosing both a blood pressure target and a norepinephrine dose for an individual…
Key concepts
- Reduced PK/PD Model
- This is a simplified mathematical model used to describe how a patient's blood pressure (MAP) responds to norepinephrine (NE) delivery. The original complex model is simplified by setting certain parameters, resulting in a smaller, linear equation that is easier to fit using only short patient data.
- Uncertainty Propagation
- This technique calculates the potential variation or risk associated with future blood pressure readings. It uses a Gaussian approximation to predict the distribution of MAP values after calibration, helping the system understand how uncertain its predictions are for subsequent measurements.
- Personalized Target Selection
- The core decision-making process finds the lowest blood pressure target that ensures a low probability (e.g., below 5%) of experiencing a reading below 65 mmHg during the critical rise phase after treatment, balancing safety with minimizing intervention.
- Control Input Generation
- This step determines the optimal way to deliver NE—specifically, how much to give initially (bolus) and how much to keep giving continuously (infusion)—to achieve the desired pressure target while minimizing the time spent at low pressure.
Summary
A risk-aware individualized blood pressure target to avoid hypotension in surgery proposes an online method for choosing both a blood pressure target and a norepinephrine dose for an individual surgical patient From a short record of mean arterial pressure (MAP) and norepinephrine delivery, the method fits a reduced pharmacokinetic/pharmacodynamic model and uses a local Gaussian approximation to propagate uncertainty to future MAP readings It searches for the lowest target for which a bolus and constant infusion keep the predicted probability of at least one MAP reading below 65 mmHg after the rise phase within a prescribed limit
Model Calibration and Simplification
The approach begins by calibrating a response model to Norepinephrine (NE) from a few minutes of Mean Arterial Pressure (MAP) readings The initial pharmacokinetic/pharmacodynamic model has twelve parameters, but this is simplified by imposing Tlag = 0 and β = 1 to create a delay-free, linear version The MAP (mmHg) depends on the total concentration C as follows: y = E0 + (Emax − E0) C / γ Cγ + ECγ / 50 Model simplification leads to a reduced model where the output is described by y = h(Ac) = Emax − a / b + Ac The seven parameters of this reduced model are kept in the rest of the paper after ranking and selection analysis
Uncertainty Propagation and Target Selection
The uncertainty is computed and propagated to future MAP readings using a first-order approximation of the distribution of future MAP values after calibration The personalized target is defined as the lowest MAP target whose control input keeps the risk of a reading below 65 mmHg at most α, or no injection when no target is safe This involves defining a forecast law, which is approximately Gaussian N (ˆyt, Σt), where Σt = S′tPtS′⊤t + σ2tI The personalized target is then found by bisection to 0.05 mmHg, assuming the risk falls as the target rises
Control Input Generation
The control input built from a bolus and an infusion is derived by inverting the model The constant infusion sets the equilibrium, and the bolus shortens the transient to minimize hypotensive time For a constant input u¯, the required input to achieve steady-state output yref is therefore u¯ = kel h−1(yref) The bolus is implemented as a rectangular pulse of duration τb = 10 s and amplitude B/τb
Application and Findings
The method was assessed on both the clinical cohort and a virtual cohort of 4,096 patients In the virtual cohort, the method recommended a dose for 74 % of patients at a nominal risk of 5 % Among these patients, 6.9 % had a reading below 65 mmHg A fixed target of 70 mmHg yielded a rate of 25 % when applied to the full virtual cohort The results suggest that y⋆ pays for the fit’s uncertainty, selecting fewer patients and setting higher targets compared to other rules
Discussion on Forecast Confidence
The forecast law is overconfident once the fit has seen enough readings, as evidenced by the median score reaching 2.13 after 15 min on the clinical cohort Overconfidence can be caused by several approximations, including treating σt as known when it is estimated from readings Model mismatch can persist across successive MAP readings if the model misses part of the patient’s response
Conclusion
We proposed an online method that fits a reduced PK/PD model to MAP readings and uses its forecast to select a personalised MAP target The method is designed to return the decision to the clinician when no target is safe, carrying the fit’s uncertainty into the dose Larger cohorts from routine intraoperative care could improve the population model The challenge remains clinical, as clinicians must keep patients above 65 mmHg during collection time The question of collection informativeness also applies to these retrospective data
Pruning redundant reasoning tokens
This paper proposes an online method for choosing both a blood pressure target and a norepinephrine dose for an individual surgical patient. The single most informative sentence about the paper's subject and finding is that it proposes an online method for choosing both a blood pressure target and a norepinephrine dose for an individual surgical patient.
How the training signal is built
The training signal, which is used to estimate the personalized target, is derived from the law of the coming MAP readings from a first-order approximation of the distribution of future MAP values after calibration This involves defining a forecast law, which is approximately Gaussian N (ˆyt, Σt), where Σt = S′tPtS′⊤t + σ2tI The personalized target is then found by bisection to 0.05 mmHg, assuming the risk falls as the target rises
Where the gains come from
The primary gain comes from individualizing a model from MAP measurements and providing dose recommendations for clinician approval This approach relies on a reduced NE PK/PD model whose seven parameters are estimated from MAP readings alone, at little cost in fit quality on the cohort of [4] The method successfully recommends a dose for 74 % of patients at a nominal risk of 5 % among the 4,096 virtual patients This demonstrates that y⋆ pays for the fit’s uncertainty by selecting fewer patients and setting higher targets The gap between the personalized target and population targets points to the tails of the forecast law, already too light on the clinical cohort The method is found to be better than an alternative designed after Joachim’s baseline in terms of infusion plan design
Collection informativeness
The collection informativeness addresses the challenge that every virtual patient receives the same NE doses during the collection time, and nothing guarantees that these are informative enough for a given patient The authors suggest excluding windows in which MAP varies too little relative to measurement noise to identify the response dynamics Synchronizing MAP and NE delivery with other physiological signals could help identify disturbances
The single most informative sentence about the paper's subject and finding is that it proposes an online method for choosing both a blood pressure target and a norepinephrine dose for an individual surgical patient.
The training signal, which is used to estimate the personalized target, is derived from the law of the coming MAP readings from a first-order approximation of the distribution of future MAP values after calibration This involves defining a forecast law, which is approximately Gaussian N (ˆyt, Σt), where Σt = S′tPtS′⊤t + σ2tI The personalized target is then found by bisection to 0.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Individualized Dose Recommendation: The improved system can propose a personalized blood pressure target based on an individual patient's historical Mean Arterial Pressure (MAP) readings, aiming for the
lowest MAP target whose NE doses keep the predicted risk of hypotension below alpha.
-
Uncertainty-Aware Decision Making: The system incorporates patient-specific uncertainty by using a posterior distribution derived from fitting a reduced pharmacokinetic/pharmacodynamic model to MAP data, allowing it to select a target that accounts for this uncertainty, as described by
the law of the coming MAP readings.
-
Online Model Calibration: The system can perform online identification of parameters in a reduced PK/PD model directly from short records of MAP readings alone, making the process
cheap
and feasible during routine surgery. -
Safe Control Input Generation: The system generates an open-loop control law that combines a constant infusion to set equilibrium and an impulsive bolus to shape the transient response, finding the
smallest admissible dose B for which the first hitting time of the target is less than Tr.
Related papers
- ViroGym: Realistic Large-Scale Benchmarks for Evaluating Viral Proteins
- CellCanon: An Agentic Cellular World Model for Prediction, Explanation, and Observation-Driven Updating
- Probing and steering biology across Boltz-1s trunk-diffusion boundary
- Trajectory Inference of Human Aging from Cross-Sectional DNA Methylation Data
- Bridging the Modality Bottleneck in Pathology MIL through Molecularly Guided Feature Adaptation
- Molecular notation can act as a shortcut in AI benchmark evaluation