Severity Classification System

• Severity classification systems assign ordered categories to observations based on seriousness, stage, and operational urgency across various domains.
• They integrate multimodal data and diverse models—from SVMs to transformers and physics-based rules—to support tailored interventions.
• This framework enables an intermediate abstraction layer that informs triage, diagnosis, maintenance, and risk management in clinical, industrial, and cybersecurity settings.

Searching arXiv for papers on severity classification systems and related methodologies. arXiv search query: severity classification system arXiv papers A severity classification system is a framework that assigns an observation to an ordered category intended to represent seriousness, stage, consequence, or operational urgency. In recent arXiv literature, the concept spans clinical staging from speech, pathology, imaging, ICU records, and Parkinson’s disease assessments; operational risk grading in cybersecurity and system logs; crash injury and incident prioritization; and non-medical severity-aware diagnosis and reliability analysis in rotating machinery and structural engineering (Javanmardi et al., 2023, Santos et al., 2023, Shojaei et al., 24 Apr 2025, Islam et al., 19 Apr 2026, Bonhomme et al., 4 Jul 2025, Leblouba et al., 16 Aug 2025). Across these domains, severity systems differ in modality and model class, but they repeatedly combine an ordered label space with rules or learned decision functions that support triage, diagnosis, maintenance, or risk-informed intervention.

1. Scope and functional role

Severity classification is typically introduced where binary detection is insufficient for downstream decisions. In dysarthria analysis, the distinction between healthy and dysarthric speech is separated from a four-class severity task comprising very low, low, medium, and high severity, because screening and severity assessment serve different clinical purposes (Javanmardi et al., 2023). A comparable distinction appears in rotor inter-turn short-circuit diagnosis, where the target is not merely faulty versus healthy but healthy, mild, moderate, and severe states, since maintenance actions depend on how far degradation has progressed (Rozhon et al., 3 Jul 2026). In structural reliability, the same logic is expressed analytically rather than through a conventional classifier: failure frequency alone is treated as incomplete, and severity is added through the Expected Failure Deficit and a calibrated five-level system from Mild to Extreme (Leblouba et al., 16 Aug 2025).

The operational role of such systems is domain-specific but structurally similar. In vulnerability management, a provisional low/medium/high/critical prediction is used to bridge the lag between disclosure and official CVSS scoring (Bonhomme et al., 4 Jul 2025). In traffic safety, KA/BC/PDO labels are used to distinguish fatal and severe injury crashes from less consequential outcomes, with emphasis on the rare KA class (Chakraborty et al., 2021). In pathology NLP, severity serves as the top-level organizing principle for a hierarchical report-understanding pipeline before detailed diagnosis is predicted (Santos et al., 2023). This suggests that severity classification often functions as an intermediate abstraction layer between raw observations and intervention policies.

2. Label semantics and problem formulations

Severity systems in the cited literature are not uniform. Some are binary, some multiclass, some hierarchical, and some explicitly ordinal. Binary formulations appear when the principal decision is escalation versus non-escalation, as in mild-to-moderate versus severe COPD in ICU data, or severe outcome within one month versus non-severe progression in COVID-19 CT-based prognosis (Shojaei et al., 24 Apr 2025, Müller et al., 2023). Multiclass formulations dominate when the application requires finer staging, such as the four-class dysarthria task, the three-class Parkinson’s disease task Healthy/Mild/Moderate-to-Severe, and the three-class crash injury target KA/BC/PDO (Javanmardi et al., 2023, Islam et al., 19 Apr 2026, Chakraborty et al., 2021). Hierarchical formulations appear when severity is first predicted and then used to constrain diagnosis branches, as in breast pathology reports (Santos et al., 2023).

The ordering of labels is frequently explicit rather than incidental. The ranking-guided domain adaptation framework for medical imaging defines severity classification as prediction over ordered labels $y \in \{1,\dots,C\}$, emphasizing that adjacent grades are clinically closer than distant ones (Harada et al., 2 Apr 2026). The hierarchical Neyman–Pearson framework for COVID-19 severity similarly assumes ordered classes, with smaller indices denoting higher priority, and formalizes “under-classification” as assigning a patient to a less severe class than the true one (Wang et al., 2022). In structural reliability, the calibrated thresholds $0.283$, $0.373$, $0.525$, and $0.7979$ divide the normalized expected deficit into five severity levels, with the highest band corresponding to a regime where the Gaussian inverse mapping is undefined and interpreted as Extreme severity (Leblouba et al., 16 Aug 2025).

Domain	Severity labels	Structure
Dysarthria speech	very low, low, medium, high	4-class acoustic severity (Javanmardi et al., 2023)
Traffic crash injury	KA, BC, PDO	3-class severity collapsed from KABCO (Chakraborty et al., 2021)
Vulnerability triage	low, medium, high, critical	4-class CVSS-derived text classification (Bonhomme et al., 4 Jul 2025)
Parkinson’s disease	Healthy, Mild PD, Moderate-to-Severe PD	3-class H&Y grouping (Islam et al., 19 Apr 2026)
Breast pathology reports	IBC, ISC, HRL, BLL, NBC, Benign, Negative	first-level severity hierarchy (Santos et al., 2023)
Structural reliability	Mild, Moderate, High, Critical, Extreme	analytically calibrated five-level system (Leblouba et al., 16 Aug 2025)

A common misconception is that severity classification is always a standard flat multiclass problem. The literature instead shows several distinct formulations: binary severity discrimination, ordinal classification, hierarchical routing, benchmark-style probing of semantic competence, and analytically derived classification from reliability quantities (Shojaei et al., 24 Apr 2025, Harada et al., 2 Apr 2026, Santos et al., 2023, Masri et al., 12 Jan 2026, Leblouba et al., 16 Aug 2025).

3. Inputs, representations, and feature construction

The input space for severity systems is highly heterogeneous. Acoustic severity classification in dysarthria uses isolated-word speech from UA-speech and compares conventional spectrogram, mel-spectrogram, and MFCC baselines with pre-trained wav2vec 2.0 embeddings extracted from the input to the first transformer block and from each of the 12 transformer blocks, averaged over non-overlapping 20 ms frames into 768-dimensional utterance-level features (Javanmardi et al., 2023). Cross-lingual dysarthria severity classification instead uses 39 handcrafted acoustic features covering voice quality, pronunciation, and prosody, then separates shared and language-specific subsets across English, Korean, and Tamil (Yeo et al., 2022).

Other systems are deliberately multimodal. The rotor ITSC framework uses a single eddy-current displacement sensor, a current probe, and an 18-dimensional hybrid feature vector comprising spectral, temporal/statistical, current-based, and operating-parameter features (Rozhon et al., 3 Jul 2026). The COVID-19 severity detector combines 3D thorax CT with age, sex, and Infection-Lung Ratio, where $$\mathrm{ILR} = \frac{|\mathrm{Infection}|}{|\mathrm{Infection}| + |\mathrm{Lung}|}$$ (Müller et al., 2023). The COPD ICU classifier uses 10 bedside variables, including blood gas measurements and vital signs, and supplements expert-derived labels with label propagation and label spreading over 3,488 initially unlabeled samples (Shojaei et al., 24 Apr 2025). STEP-PD uses 208 multimodal clinical features from subjective questionnaires and clinician-assessed measures across 15,606 visit-level samples (Islam et al., 19 Apr 2026).

Text-centric severity systems convert unstructured language into severity signals. VLAI uses the vulnerability description alone, tokenized with RoBERTa’s byte-pair tokenizer and truncated or padded to 512 tokens (Bonhomme et al., 4 Jul 2025). HCSBC applies PathologyBERT-Base Uncased to breast pathology reports, treating severity as the first-stage target before diagnosis-level prediction (Santos et al., 2023). The traffic incident system converts each record into a full-text representation of column name: column value plus the narrative description, then extracts dense language-model features that are combined with structured report features (Grigorev et al., 2024). By contrast, the proposal on production log severity classification leaves the final feature set open, but frames the task as stream processing over production log messages and metadata (Mendes et al., 2021).

These representation choices are often physically or clinically motivated rather than purely empirical. The ITSC study chooses the $4f_r$ harmonic as a main diagnostic carrier because it is more sensitive to fault progression than the dominant $3f_r$ component (Rozhon et al., 3 Jul 2026). The dysarthria wav2vec study interprets earlier layers as preserving more generic acoustic information and later layers as more linguistic or phoneme-identity oriented, which aligns with its observed early-layer advantage for detection and late-layer advantage for severity (Javanmardi et al., 2023). The reliability framework defines severity through the normalized expected failure deficit $E_f^\ast = E_f/\sigma_g$, making severity a property of failure depth rather than only failure frequency (Leblouba et al., 16 Aug 2025).

4. Model architectures and decision mechanisms

Severity classification systems use a wide range of decision mechanisms, from simple discriminative classifiers to hierarchical control rules. In the dysarthria wav2vec study, a binary SVM and a one-vs-one SVM with RBF kernel are trained on utterance-level features, with $\gamma = \frac{1}{D \cdot Var(X)}$ and z-score normalization computed from training data (Javanmardi et al., 2023). In ICU COPD severity classification, random forest, KNN, and SVM are compared after semi-supervised labeling, with random forest emerging as the strongest model (Shojaei et al., 24 Apr 2025). In crash injury severity, decision tree, random forest, XGBoost, and a DNN with four hidden layers of 128, 128, 128, and 64 neurons are trained after Granger-causality-based feature selection and imbalance handling via undersampling and SMOTE (Chakraborty et al., 2021).

Transformer fine-tuning dominates text-only severity tasks. VLAI uses RoBERTa-base with a softmax head and the decision rule $0.283$0, trained with cross-entropy, AdamW, and a linear learning-rate scheduler with warm-up (Bonhomme et al., 4 Jul 2025). HCSBC uses a two-level architecture in which PathologyBERT first predicts one or more of seven severity categories, after which the report is routed to one of six diagnosis branches, with weighted binary cross entropy to address class imbalance (Santos et al., 2023). STEP-PD benchmarks Logistic Regression, SVM, KNN, Random Forest, and XGBoost, then uses SHAP for both global and local explanations once XGBoost is selected as the strongest model (Islam et al., 19 Apr 2026).

Several papers depart from a single-model paradigm. OASIC first estimates occlusion severity at test time from the anomaly map average $0.283$1, masks the occluder with gray fill, and then selects the classifier trained for the nearest occlusion range (Gijzen et al., 5 Apr 2026). The hierarchical Neyman–Pearson framework constructs ordered thresholds over class scores so that severe classes are tested first and under-classification errors are controlled with high probability (Wang et al., 2022). The ranking-guided semi-supervised domain adaptation method introduces a classifier head $0.283$2 and ranking head $0.283$3, then aligns source and target domains using Cross-Domain Ranking and Continuous Distribution Alignment rather than ordinary class-wise clustering (Harada et al., 2 Apr 2026). In structural reliability, the classifier-like object is an inverse map from normalized expected failure deficit to a severity-aware reliability index $0.283$4, defined by $0.283$5 when the inverse exists (Leblouba et al., 16 Aug 2025).

A second misconception is that severity systems are necessarily end-to-end black boxes. The cited work includes end-to-end transformers and deep CNNs, but also physics-based feature systems, hybrid pipelines, rule-based clinical labeling, and analytical calibrations (Santos et al., 2023, Rozhon et al., 3 Jul 2026, Shojaei et al., 24 Apr 2025, Leblouba et al., 16 Aug 2025).

5. Evaluation criteria and empirical behavior

Evaluation protocols vary with task structure. The dysarthria severity system uses 81 balanced iterations with one speaker from each class held out per iteration, whereas dysarthria detection uses leave-one-speaker-out cross-validation across 28 speakers (Javanmardi et al., 2023). COPD severity uses stratified 5-fold cross-validation and reports accuracy, precision, recall, F1, and ROC AUC (Shojaei et al., 24 Apr 2025). VLAI combines a held-out test set with a live retrospective evaluation on vulnerabilities that were initially unscored (Bonhomme et al., 4 Jul 2025). OASIC evaluates robustness by $0.283$6, integrating top-1 accuracy across occlusion severities, while crash injury modeling emphasizes normalized confusion matrices because overall accuracy alone is misleading under extreme imbalance (Gijzen et al., 5 Apr 2026, Chakraborty et al., 2021).

Representative reported results illustrate both domain dependence and the importance of formulation choice.

System	Task	Reported result
wav2vec dysarthria severity	4-class severity	wav2vec-13: 44.56% accuracy; +10.62% absolute over MFCCs (Javanmardi et al., 2023)
VLAI	4-class vulnerability severity	82.8% test accuracy; about 85% live retrospective agreement (Bonhomme et al., 4 Jul 2025)
Rotor ITSC severity	4-class fault severity	90.56% accuracy; healthy recall 0.99; mild recall 0.87 (Rozhon et al., 3 Jul 2026)
COPD ICU severity	binary mild-to-moderate vs severe	Random Forest: 0.9251 ± 0.0105 accuracy; 0.9841 ± 0.0030 ROC AUC (Shojaei et al., 24 Apr 2025)
STEP-PD	3-class PD severity	XGBoost: 0.9414 ± 0.0050 accuracy; 0.8775 ± 0.0145 Macro-F1 (Islam et al., 19 Apr 2026)
OASIC	occlusion-severity-informed classification	$0.283$7 (Gijzen et al., 5 Apr 2026)

Other reported outcomes reinforce recurring patterns. In dysarthria, the earliest wav2vec layer is best for binary detection at 93.95% accuracy, whereas the final layers are best for severity, supporting a task-dependent layer utility pattern (Javanmardi et al., 2023). In HCSBC, the PathologyBERT-based severity classifier achieves about 0.93 accuracy and about 0.93 Micro F1 internally, while external validation remains around 84% on MGH and 86% on Mayo after training on EUH (Santos et al., 2023). Cross-lingual dysarthria severity reaches a 67.14% F1 score, improving over both Intersection and Union baselines and exceeding monolingual classification for English, Korean, and Tamil (Yeo et al., 2022). In log comprehension benchmarking, severity classification accuracy ranges from 95.64% for Qwen3-4B with RAG to 0.00% for Phi-4-Mini-Reasoning under both few-shot and RAG prompting, indicating that performance depends strongly on retrieval integration under strict output constraints (Masri et al., 12 Jan 2026).

Class-wise behavior is frequently more informative than global scores. The crash injury study reports that decision tree is best for PDO, random forest for BC, and DNN for the rare KA class, rather than identifying a single universally dominant classifier (Chakraborty et al., 2021). The wound-image study reaches 68.49% multi-class accuracy with VGG19, but binary tasks are easier, achieving 78.79% for green versus yellow, 81.40% for green versus red, and 77.57% for yellow versus red (Anisuzzaman et al., 2022). In coffee leaf diagnosis, severity estimation remains harder than stress-type classification, and errors tend to stay near the diagonal of the confusion matrix because the severity labels are ordinal (Esgario et al., 2019).

6. Limitations, controversies, and current directions

Several limitations recur across the literature. Class imbalance is persistent: PDO crashes are more than 30 times as frequent as KA crashes in the Texas dataset, and the Mod-Severe class in STEP-PD contains only 553 visit-level samples out of 15,606 (Chakraborty et al., 2021, Islam et al., 19 Apr 2026). Borderline class boundaries are another recurring issue. The ranking-guided domain adaptation paper argues that severity classes are continuous in nature and therefore poorly matched to assumptions of sharply separated clusters (Harada et al., 2 Apr 2026). The breast pathology study notes high linguistic variability, multiple diagnoses per report, and underrepresented rare subtypes (Santos et al., 2023). The wound-severity study states that factors such as size, depth, callus tissue, maceration, and breakdown are not fully captured by 2D image convolution alone (Anisuzzaman et al., 2022).

Label quality is also contested. The production-log proposal argues that severity assignment is subjective and influenced by developer experience and preferences, while the later system-log benchmark goes further by treating severity prediction as a probe for runtime log comprehension rather than an end task, because Syslog severity labels are predefined metadata, administrator-defined, and weakly standardized (Mendes et al., 2021, Masri et al., 12 Jan 2026). VLAI likewise positions its output as supplementary and provisional rather than authoritative, and explicitly notes dependence on the wording of textual descriptions and vulnerability to strategic omission of phrases such as “remote code execution” (Bonhomme et al., 4 Jul 2025). In COPD severity classification, the initial labels are based on pulmonologist-informed rules, which the paper identifies as a source of subjectivity (Shojaei et al., 24 Apr 2025).

Generalization remains a central challenge. The wav2vec dysarthria study is limited to UA-speech and calls for cross-database evaluation and assessment on other disorders (Javanmardi et al., 2023). The rotor ITSC framework is validated on a laboratory test rig and still requires testing across different ratings and real industrial environments (Rozhon et al., 3 Jul 2026). The COVID-19 CT severity system explicitly states that additional validation and integration into a clinical study are required to establish real clinical benefit (Müller et al., 2023). The traffic-incident LLM hybrid reports strong gains on U.S. data but mixed improvements on U.K. and Queensland datasets, indicating that transfer across reporting conventions is nontrivial (Grigorev et al., 2024).

Recent work also shows a shift from accuracy-only severity classification toward risk-aware and explanation-aware frameworks. The hierarchical Neyman–Pearson method formalizes asymmetric penalties for missing severe cases and supplies high-probability control of under-classification errors (Wang et al., 2022). STEP-PD uses SHAP to show a stage-dependent change from early motor features to axial and balance impairments (Islam et al., 19 Apr 2026). The structural reliability framework adds severity to classical reliability by measuring conditional failure depth rather than only failure probability (Leblouba et al., 16 Aug 2025). A plausible implication is that future severity systems will increasingly combine ordered label modeling, error asymmetry, interpretability, and external validation rather than treating severity as a conventional multiclass benchmark alone.