Time-Critical Reasoning: Representations and Application

Eric Horvitz, Adam Seiver

Introduction

In high-stakes, time-sensitive environments such as trauma surgery or space shuttle operations, the value of an action is inextricably linked to when it is performed. Traditionally, decision-making in these "time-critical" contexts has relied on complex temporal models like Partially Observable Markov Decision Processes (POMDPs) or dynamic Bayesian networks. While mathematically robust, these models often impose a computational burden that is impractical during emergencies. This paper proposes a simplified framework focusing on time-dependent utility: quantifying the loss of value—or life—for every minute an action is delayed.

Methodology: Reformulating the Decision Problem

The authors identify a class of "pathological-process problems" where a deleterious process (e.g., internal bleeding) is underway. Rather than tracking every micro-step of the system's evolution, they reformulate the problem to assess outcomes at a "fixed horizon" or "equilibrium."

Key Metrics

  • Time-Dependent Utility ($u(A, H, t)$): This function maps an action ($A$), a hypothesis about the system state ($H$), and the duration of the process ($t$) directly to a utility score.
  • Expected Cost of Delayed Action (ECDA): This metric calculates the difference in expected utility between acting immediately versus waiting until a later time $t$. It provides a concrete measure of urgency: "Waiting 10 minutes will cost X units of expected survival."
  • Bayesian Criticality-Assessment: To feed these utility curves, the authors employ Bayesian networks to infer the probability of various physiological syndromes from observed symptoms.
A Bayesian network used to infer physiological criticality from observed signs and symptoms.

Application and Results

The framework was applied to two primary domains: trauma care and NASA's propulsion systems.

Trauma Triage

Using expert-elicited knowledge from trauma surgeons, the authors developed survival curves for various injuries. These curves demonstrate how the probability of survival ($Pr(Survive)$) decays non-linearly over time. For instance, a major airway obstruction requires immediate intervention (high ECDA), whereas a minor fracture has a much flatter utility curve.

Survival curves for various trauma injuries illustrate the time-dependent utility of medical intervention.

NASA Vista System

In the context of the Space Shuttle, the method allowed ground controllers to prioritize telemetry faults based on "criticality." By focusing on the cost of delay rather than just the likelihood of a fault, controllers could address the most time-sensitive failures first.

Key Insights

  • Urgency vs. Probability: A highly probable condition (e.g., a permanent brain injury) may not be urgent if its outcome is not sensitive to timely action. Conversely, a less likely but rapidly fatal condition (e.g., a tension pneumothorax) demands immediate attention due to its high ECDA.
  • The "Life Lottery": By assessing utility based on the "future life lottery" at equilibrium, the model simplifies complex temporal dependencies into a single outcome assessment.
  • Load-and-Go vs. Stay-and-Play: The ECDA provides a formal mathematical framework for deciding whether to stabilize a patient on-site or transport them immediately, by calculating the "equivalent time removed" by local treatment.

Limitations and Future Work

The current model assumes a primary pathology, whereas real-world cases often involve multiple interacting conditions. Additionally, determining "Time Zero"—the exact start of a pathological process—remains a challenge. Future research aims to integrate real-time environmental data, such as traffic conditions, to dynamically update transport time variables and refine the modeling of suboptimal future actions.

Summary updated

Abstract

We review the problem of time-critical action and discuss a reformulation that shifts knowledge acquisition from the assessment of complex temporal probabilistic dependencies to the direct assessment of time-dependent utilities over key outcomes of interest. We dwell on a class of decision problems characterized by the centrality of diagnosing and reacting in a timely manner to pathological processes. We motivate key ideas in the context of trauma-care triage and transportation decisions.
BibTeX
@inproceedings{horvitz1997time,
  author = {Eric Horvitz and Adam Seiver},
  title = {{Time-Critical Reasoning: Representations and Application}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {1997},
  url = {https://openreview.net/forum?id=E3FhQVZw-kh},
}
Figure