Criticality Score: Definition & Applications

• Criticality Score is a quantitative metric that measures system importance, risk, and vulnerability in diverse technical domains using tailored mathematical formulations.
• It integrates methodologies from reinforcement learning, traffic safety, infrastructure planning, neural modeling, and business continuity to aid decision-making and resource allocation.
• Empirical outcomes demonstrate that calibrated criticality scores enhance efficiency and risk mitigation, proving vital in safety-critical and complex operational environments.

A criticality score is a quantitative metric or function designed to measure the importance, hazard, vulnerability, or potential impact associated with a particular system state, entity, or scenario across diverse technical domains. The criticality score concept appears in reinforcement learning (RL), risk analysis for automated driving, infrastructure resilience planning, neurosystems modeling, spiking neural networks (SNNs), business continuity management, intensive care medicine, and nuclear safety. While its concrete definitions are domain-specific, criticality scores generally serve as decision-making aids for prioritization, intervention, risk assessment, or resource allocation.

1. Formal Definitions and Representative Mathematical Formulations

Criticality scores are instantiated with precise mathematical formulations tailored to each field:

• Reinforcement Learning (Policy Sensitivity):

A criticality function $h : S \to [0,1]$ assigns to each MDP state $s$ a scalar denoting the expected influence of action choice on cumulative reward, often approximated by the variance of the optimal action-value function: $h(s) \propto \mathrm{Var}_a[Q^*(s,a)]$ (Spielberg et al., 2018). Alternatively, the true criticality at time $t$ for $n$-step deviation is defined as

$$c(t, n; \pi) = \mathbb{E}_{a\sim\pi}[R_\gamma] - \mathbb{E}_{a\sim\pi'(t,n)}[R_\gamma],$$

where $\pi'(t, n)$ is a perturbed policy taking random actions for $n$ steps at $t$ (Grushin et al., 2024).

• Automated Driving and Risk Analysis:

Criticality metrics are based on physical models such as Time-to-Collision (TTC), minimum safety margins, and surrogate deceleration measures. For complex, crowded scenes, composite metrics like the Inverse Universal Traffic Quality (IUTQ) combine macroscopic (system variability), metascopic (proximity), mesoscopic (local speed variance), and microscopic (ego dynamics) sub-scores via

$$\mathrm{TQ}_{\mathrm{co}} = \sqrt{ \mathrm{TQ}_\Omega^2 + \mathrm{TQ}_\eta^2 + \mathrm{TQ}_\theta^2 + \mathrm{TQ}_\mu^2 }$$

with criticality thresholds calibrated empirically (Schütt et al., 2023, Westhofen et al., 2021). Object-specific criticality labels for perception systems rely on time-based and distance-based metrics aggregated via logical operations (multi-metric aggregation, bidirectional OR) (Gamerdinger et al., 17 Dec 2025).

• Infrastructure and Facility Dependence:

Functional criticality of facility $f$ is computed as

$$\mathrm{FC}_f = \frac{1}{N_f}\sum_{i=1}^{N_f} \frac{V_{i,f}}{s_i},$$

where $V_{i,f}$ is the number of visits from origin $i$, $N_f$ is the number of distinct origins, and $s_i$ is the substitutability (count of adjacent alternatives). The normalized criticality score is

$$C_f = \frac{\mathrm{FC}_f - \min_g\mathrm{FC}_g}{\max_g\mathrm{FC}_g - \min_g\mathrm{FC}_g}, \quad 0 \le C_f \le 1$$

(Ma et al., 18 Dec 2025).

• Neural Circuit and SNN Criticality:

The criticality score for a neuron $e$ is defined as

$$C(e) = \frac{1}{N}\sum_{i=1}^{N} \mathrm{Aggregate}_t\left( \frac{1}{T} \sum_{t=1}^T g'(u_{e,t}(x_i)) \right),$$

where $g'(u)$ is the derivative of the surrogate activation, measuring proximity to a critical membrane threshold (Chen et al., 2023).

• Business Continuity and IT Risk:

The Business Continuity Testing Points (BCTP) method defines the criticality score as

$$\mathrm{ABFRP} = \mathrm{UBFRP} \times \mathrm{TRF} \times \mathrm{ERF} \times \mathrm{URF}$$

where UBFRP is an unadjusted sum over actor and process weights, and TRF/ERF/URF are technical, environmental, and unexpected recovery factors (Podaras et al., 2013).

2. Domain-Specific Use Cases and Interpretations

Criticality scores are consistently used to prioritize, warn, or optimize according to the risk, leverage, or dependency embodied by a system element or scenario:

• RL/Autonomous Agents:

Criticality identifies states or times where policy deviations have disproportionate impact, guiding attention, human oversight, or adaptive algorithm steps (Spielberg et al., 2018, Grushin et al., 2024). For example, safety margins derived from true or proxy criticality provide real-time thresholds for intervention (Grushin et al., 2024).

• Traffic Safety and Automated Driving:

Criticality metrics filter or rank dangerous situations, score relevance of detections, or assist in scenario generation for validation (Westhofen et al., 2021, Gamerdinger et al., 17 Dec 2025). Aggregate metrics (e.g., IUTQ) provide holistic scene-level scores, while pairwise or multi-metric criticality enables robust identification of safety-relevant objects in complex interactions (Schütt et al., 2023, Gamerdinger et al., 17 Dec 2025).

• Infrastructure Resilience:

Behavioral functional criticality exposes which nodes (e.g., specific hospitals or grocery stores) represent irreplaceable lifelines. Coupling criticality with hazard models (e.g., probabilistic flood exposure) yields population-weighted risk multipliers (Ma et al., 18 Dec 2025).

• Neural Systems and SNNs:

In neuroscientific models or SNN pruning, criticality scores signal maximally information-rich components—either temporal regimes (cortical phase transition) (Kozma et al., 2012) or neurons most vital for feature entropy (Chen et al., 2023).

• Business Continuity:

Quantitative criticality guides assignment of IT functions to correct recovery test categories, reflecting risk to operations under crisis (Podaras et al., 2013).

• Clinical Risk and ICU Prognosis:

Continuous criticalness scores augment discrete risk classes, enabling regression-based risk stratification aligned with observed clinical outcomes (Sahu et al., 1 Aug 2025, Arzeno et al., 2014).

3. Methodologies for Calculation and Calibration

The computation of a criticality score involves domain-appropriate workflows:

• RL:
  • For policy deviation, sample rollouts of baseline and perturbed policies, compute empirical criticality $c^*(t,n)$, then fit statistically monotonic proxies for efficient run-time deployment (Grushin et al., 2024).
  • In CVS-style frameworks, cumulative human-provided criticality per state determines adaptive algorithmic steps (Spielberg et al., 2018).
• Safety and Traffic:
  • Compute physical/dynamical metrics (e.g. TTC, PET, surrogate deceleration) per object, threshold to obtain binary relevance, then fuse directionally or across metrics. Aggregate scene metrics like IUTQ require calculation across all actors and sub-metrics per frame (Westhofen et al., 2021, Gamerdinger et al., 17 Dec 2025, Schütt et al., 2023).
• Infrastructure:
  • Aggregate human-origin visitation statistics, normalize by substitutability, and rescale across entire facility portfolios. Propagate scores into exposure- or vulnerability-weighted regional indices (Ma et al., 18 Dec 2025).
• Neuroscience/SNNs:
  • For SNNs, aggregate the surrogate gradient at each neuron temporally and over batch data. In neuropercolation, calculate higher-order cumulants (e.g., Binder cumulant $u_4$) as a function of network parameters, and locate cross-scale invariances that index critical regime proximity (Kozma et al., 2012, Chen et al., 2023).
• Business Continuity:
  • Enumerate actors and processes, apply domain-calibrated weights, then multiplicatively adjust for recovery factors (Podaras et al., 2013).
• Clinical Scores:
  • Fit logistic surrogate functions per variable, then optimize smooth weights and thresholds by regularized likelihood or MSE calibration against observed outcomes; update periodically for clinical drift (Arzeno et al., 2014, Sahu et al., 1 Aug 2025).

4. Impact, Applications, and Empirical Outcomes

Criticality score deployment demonstrates quantifiable benefits:

• RL and Agent Oversight:

Adaptive step computation via criticality achieves marked sample-efficiency improvements over baseline algorithms (Spielberg et al., 2018). Safety-margin coverage demonstrates that selective human monitoring at the 5% most critical points can intercept nearly half of catastrophic outcomes (Grushin et al., 2024).

• Automated Driving and Perception Evaluation:

Scene-level metrics like IUTQ outperform traditional TTC/PET/F-score measures for crowded urban environments (MCC = 0.654 vs TTC 0.550) (Schütt et al., 2023). In object relevance classification, bidirectional and multi-metric criticality increase framewise recall up to 100% compared to single-metric baselines (Gamerdinger et al., 17 Dec 2025). Suitability analysis frameworks help practitioners select metrics tuned to application needs (Westhofen et al., 2021).

• Infrastructure Planning:

Functional criticality reveals highly concentrated facility-dependence: 2.8% of grocery stores (resp. 14.8% hospitals) serve disproportionate demand, with coupling to flood models showing vulnerability increases outpace hazard increases (~67% rise in population-weighted risk despite modest physical hazard change) (Ma et al., 18 Dec 2025).

• SNNs and Neural Modeling:

Incorporation of neuron-level criticality in pruning—via low-cost, information-theoretic surrogates—reduces resource requirements by up to 95% and improves post-prune accuracy vs magnitude-based baselines (Chen et al., 2023). In cortical criticality models, distance-to-criticality scores enable the identification of physiologically optimal dynamical regimes (Kozma et al., 2012).

• Clinical and Business Risk:

Smooth, data-fitted criticality/prognostic scores for ICU patients yield higher discriminative power (AUC improvement of 2–21% beyond stepwise benchmarks, with superior calibration) and provide granular risk stratification for intervention (Arzeno et al., 2014, Sahu et al., 1 Aug 2025). Business process criticality enables rational triage of continuity exercises without arbitrary or purely experiential assignment (Podaras et al., 2013).

5. Limitations, Pitfalls, and Practical Recommendations

Domain-specific limitations include:

• Proxy criticality in RL can be heuristic; offline calibration is needed to ensure monotonicity and statistical confidence (Grushin et al., 2024).
• Safety metrics for driving are sensitive to modeling assumptions (e.g., constant velocity in TTC); combining complementary metrics and bi-directional logic is essential for robustness (Westhofen et al., 2021, Gamerdinger et al., 17 Dec 2025).
• Infrastructure criticality is sensitive to the definition of substitutability and spatiotemporal stationarity of mobility data (Ma et al., 18 Dec 2025).
• SNN and neural-score aggregation may miss dynamical synergies if only per-neuron scalar averages are used (Chen et al., 2023).
• Business and clinical scores require periodic recalibration as system/environmental characteristics drift; adoption depends on the availability of high-quality, regularly updated data (Arzeno et al., 2014, Podaras et al., 2013).

Recommendations include: calibrate and validate criticality scores using ground-truth or observed adverse event coverage; apply multi-metric and scenario-adaptive methodology; revisit scoring procedures as systems or environments change.

6. Comparative Framework and Cross-Domain Synthesis

Despite formulaic diversity, criticality scores across domains share several unifying themes:

• They convert complex, multidimensional, or time-varying input (state, configuration, behavior, encounter geometry, flows) into a summary value representing leverage, risk, or dependence.
• Their practical value lies in triage—prioritizing intervention (human or automated), monitoring, simulation, or resource allocation.
• Calibration is empirical, leveraging monotonic proxy relationships, expert rules, domain data, or optimization.
• In safety-critical systems (nuclear, clinical, traffic, RL agents), criticality scoring directly supports threshold-based operational constraints and oversight (e.g., nuclear $\leq 0.95$ $k_\mathrm{eff}$ requirements (Hutchinson et al., 2021)).

The cross-domain use of criticality scores reflects the growing need for interpretable, actionable, and data-linked prioritization metrics in increasingly autonomous, interconnected, and high-dimensional operational environments.