📊 statistics

Identification of physiological shock in intensive care units via Bayesian regime switching models

This paper proposes a Bayesian regime switching model that analyzes longitudinal vital signs and lab data from a large Mayo Clinic dataset to probabilistically detect occult hemorrhage and physiological shock in ICU patients, thereby enabling earlier clinical intervention.

Original authors: Emmett B. Kendall, Jonathan P. Williams, Curtis B. Storlie, Misty A. Radosevich, Erica D. Wittwer, Matthew A. Warner

Published 2026-03-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Emmett B. Kendall, Jonathan P. Williams, Curtis B. Storlie, Misty A. Radosevich, Erica D. Wittwer, Matthew A. Warner

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a crime, but the criminal is internal bleeding inside a patient's body. The problem? The crime scene is hidden. The patient might look fine on the outside, but inside, they are losing blood. By the time the "smoke" (obvious symptoms like pale skin or a very fast heartbeat) becomes visible, it might be too late to save them.

This paper is about building a super-smart, digital detective that can spot this hidden bleeding before it becomes a catastrophe.

Here is how the researchers built this detective, explained in simple terms:

1. The Problem: The "Silent Killer"

In the Intensive Care Unit (ICU), doctors watch patients closely. They look at vital signs like heart rate, blood pressure, and blood tests. But sometimes, a patient starts bleeding internally, and their body is so good at compensating that the vital signs don't change immediately. It's like a car with a slow leak in the tire; the car drives fine for a while, but suddenly, the tire blows out.

The researchers wanted a system that notices the tiny changes in the tire pressure long before the blowout happens.

2. The Solution: A "Shape-Shifting" Model

The team created a computer model called a Bayesian Regime Switching Model. That's a fancy name for a system that understands that patients aren't static; they are constantly shifting between different "modes" or "states."

Think of a patient's health like the weather.

State 1 (Sunny): The patient is stable and healthy.
State 2 (Stormy): The patient is actively bleeding (the storm is raging).
State 3 (Clearing Up): The bleeding has stopped, and the patient is recovering.
State 4 & 5 (Cloudy/Rainy): The patient is sick or stressed, but not bleeding (maybe they have an infection or are just reacting to surgery).

The model's job is to look at the data (heart rate, blood pressure, etc.) and guess: "Is it Sunny right now, or has a Storm just started?"

3. How the Detective Works: The "Black Box" vs. The "Glass House"

Many modern AI tools are "Black Boxes." You put data in, and they give you an answer, but you have no idea how they got there. Doctors don't trust Black Boxes because they can't explain the reasoning.

This new model is a "Glass House."

It looks at four key clues: Heart Rate, Blood Pressure, Hemoglobin (blood count), and Lactate (a chemical that builds up when organs aren't getting enough oxygen).
It also checks the Medication Log. If a patient gets a drug that raises their heart rate, the model knows, "Oh, the heart rate went up because of the pill, not because of bleeding." It separates the noise from the signal.
It even accounts for what happened before the patient walked into the ICU. If they arrived already stressed or injured, the model adjusts its expectations so it doesn't get confused.

4. The Secret Sauce: The "Time-Traveling" Algorithm

The hardest part of this job is that the clues (vital signs) at 2:00 PM depend on what happened at 1:45 PM, 1:30 PM, and so on. It's a chain reaction.

Standard computer models struggle with this because the math gets too heavy, like trying to solve a Rubik's cube while juggling. The researchers invented a new, faster way to solve the puzzle.

Imagine trying to guess the path of a hiker in a forest.

Old way: Guess every single step the hiker took, one by one, checking every possible path. This takes forever.
Their new way: They look at the terrain and the hiker's likely speed to guess a block of steps at once. If the guess looks right, they keep it; if not, they try a different block. This is much faster and finds the correct path (the bleeding event) much more accurately.

5. The Test Run: Real-Life Cases

The team tested their model on 33,924 real patient records from the Mayo Clinic. They didn't just look at the numbers; they had real doctors review the results.

The Success: In one case, a patient had a hidden bleed. The model spotted the "Stormy" state hours before the doctors realized the patient was crashing. It gave a "High Probability of Bleeding" alert, allowing doctors to intervene early.
The Challenge: Sometimes, a patient is just very sick (sepsis) and their vitals look like bleeding, even though they aren't. The model sometimes gets confused here, thinking it's a storm when it's just heavy clouds. The researchers admit this is a work in progress, but the model is already better than just watching the numbers with the naked eye.

The Bottom Line

This paper isn't just about math; it's about saving time. In trauma and bleeding, the "Golden Hour" is the difference between life and death.

This model acts like a 24/7 co-pilot for ICU doctors. It doesn't replace the doctor; instead, it whispers in their ear: "Hey, look at this patient. Their numbers are shifting in a way that usually means hidden bleeding. Let's check them now, before it's too late."

By turning complex, hidden data into a simple probability ("There is a 90% chance this patient is bleeding"), the researchers hope to stop internal bleeding from being a silent killer.

1. Problem Statement

The paper addresses the critical challenge of detecting occult hemorrhage (internal bleeding) in Intensive Care Unit (ICU) patients.

Clinical Context: Internal bleeding is often not clinically apparent until significant blood loss has occurred. Delayed diagnosis leads to catastrophic outcomes, including physiological shock and death.
Data Challenges:
- Latency: Physiological compensation mechanisms often mask early signs of bleeding; vital sign deterioration (e.g., heart rate increase, MAP decrease) typically occurs only after significant blood volume loss (15–40%).
- Data Complexity: Electronic Health Records (EHR) contain high-dimensional, longitudinal data with missing values, irregular sampling, and confounding factors (e.g., medication administration).
- Label Scarcity: There is a lack of "gold-standard" labeled data for bleeding events, making supervised learning difficult.
- Confounding: Medications (vasopressors, sedatives) significantly alter vital sign trends, potentially mimicking or masking shock states.

2. Methodology

The authors propose a Bayesian Regime Switching Model (RSM), a type of state-space model, designed to infer latent physiological states from observed vital signs and lab values.

A. Model Architecture

The model defines a 5-state latent Markov process representing the patient's physiological condition:

Stable: Low risk of shock/bleeding.
Hemorrhage: Active internal bleeding.
Recovery from Hemorrhage: Post-bleeding stabilization.
Non-Bleeding Event (NBE): Physiological instability not caused by bleeding (e.g., sepsis, cardiac issues).
NBE Recovery: Stabilization from non-bleeding events.

B. Conditional Response Model

Unlike standard Hidden Markov Models (HMMs) where observations are conditionally independent given the state, this model uses an Autoregressive (AR) structure:

Observations ( $Y$ ): A 4-dimensional vector comprising Hemoglobin, Heart Rate, Mean Arterial Pressure (MAP), and Lactate.
Dependence: The observation at time $k$ depends on the entire history of latent states up to time $k$ , not just the current state. This captures the cumulative physiological impact of a regime change.
Covariates: The model explicitly incorporates:
- Medications: Modeled as covariates affecting Heart Rate and MAP to disentangle drug effects from physiological shock.
- RBC Transfusions: Used as a partial proxy for bleeding severity.
- Pre-ICU History: A novel mechanism accounts for physiological changes occurring before ICU admission (latent time spent in states 2–5 prior to the first observation).

C. Bayesian Inference & Sampling

Semi-Supervised Learning: Since explicit bleeding labels are missing, the model uses a partial-labeling scheme: if a patient receives $\ge$ 3 RBC transfusions in a 12-hour window, a bleeding event is assumed to have occurred.
MCMC Algorithm: The authors developed a novel Metropolis-within-Gibbs sampling routine.
- Challenge: Standard Gibbs sampling is computationally infeasible due to the complex dependence structure (observation at $t$ depends on all previous states).
- Solution: A novel state-sampling algorithm (Algorithm 1) that proposes state sequences over blocks of time ( $p$ ) using a forward-backward logic. This achieves $O(5(p-2) + 5^2)$ complexity, significantly outperforming standard approaches ( $O(5^p)$ ) in both speed and accuracy.
Interpretability: The model provides a discrete posterior probability distribution for the latent state at every time point, offering a transparent, probabilistic assessment rather than a "black-box" prediction.

3. Key Contributions

Novel Statistical Framework: Development of a Bayesian RSM that extends beyond standard HMMs by modeling the response as an autoregressive process dependent on the full history of latent states.
Computational Innovation: Creation of an efficient state-sampling algorithm that handles high-dimensional parameter spaces and complex dependencies, solving the computational bottleneck of traditional methods.
Clinical Integration:
- Explicit modeling of medication effects to prevent confounding.
- A mechanism to account for pre-ICU physiological changes, acknowledging that patients often arrive in unstable states.
- Use of RBC transfusion data as a semi-supervised signal to train the model without full manual annotation.
Interpretability: Unlike deep learning approaches, this model yields interpretable parameters (e.g., specific slopes for heart rate changes during hemorrhage) and provides probabilistic alerts clinicians can act upon.

4. Results

Simulation Study

Performance: The model achieved a median AUC of 0.84 in identifying the "Hemorrhage" state (State 2).
Threshold Calibration: A posterior probability threshold of 0.0465 was identified to balance sensitivity (72.7%) and specificity (96.2%).
Efficiency: The novel state-sampling algorithm was 36,000 times faster than standard Gibbs sampling approaches while achieving higher state-identification accuracy (87.66% vs. ~83%).

Real Data Analysis (Mayo Clinic)

Dataset: 33,924 ICU patient encounters (500 used for training, 5 for detailed case studies).
Parameter Estimates:
- Hemorrhage Phenotype: The model correctly identified that State 2 corresponds to decreasing Hemoglobin, increasing Heart Rate, decreasing MAP, and increasing Lactate.
- Autocorrelation: Vital signs (HR, MAP) showed high autocorrelation in the "Stable" state but low autocorrelation during shock states, reflecting physiological instability.
Case Studies:
- True Positives: Successfully detected occult bleeding in patients with confirmed hemorrhagic events (e.g., Subject C with SMA dissection, Subject D with intraoperative bleeding).
- False Positives/Nuance: The model occasionally flagged "Hemorrhage" (State 2) during septic or cardiogenic shock due to overlapping vital sign trends (tachycardia, hypotension). However, the authors argue this is clinically useful as it flags any shock state requiring intervention, even if the specific etiology requires further clinical differentiation.

5. Significance and Future Work

Clinical Impact: The model offers a tool for early warning, potentially detecting bleeding before physiological compensation fails. This aligns with the "Golden Hour" concept in trauma care, where early intervention drastically improves survival.
Resource Optimization: Early detection could reduce total transfusion requirements and prevent organ failure.
Future Directions:
- Development of a real-time R Shiny application for clinical deployment.
- Extending the model to detect other adverse events (e.g., sepsis, cardiac arrest) beyond hemorrhage.
- Refining the model to allow for multiple distinct hemorrhage events within a single patient encounter, rather than assuming a single phenotype for all bleeding events.

In summary, this paper presents a rigorous, interpretable, and computationally efficient Bayesian framework that bridges the gap between complex statistical modeling and clinical decision-making, offering a promising solution for the early detection of life-threatening physiological shock in the ICU.