An auditable pipeline for fuzzy full-text screening in systematic reviews: integrating contrastive semantic highlighting and LLM judgment Authors: Arya Rahgozar, Pouria Mortezaagha Published: 2026 ABSTRACT AbstractFull-text screening is the major bottleneck of systematic reviews (SRs), particularly in domains such as population health modelling of noncommunicable diseases (NCDs), where decisive eligibility information is scattered across long and heterogeneous full texts. In this methodological study, we introduce a scalable and auditable pipeline that reframes inclusion and exclusion as a fuzzy decision process. We evaluate the approach within the Population Health Modelling Consensus Reporting Network (POPCORN) and benchmark it against statistical and crisp baselines. Articles are parsed into overlapping chunks and embedded with a domain-adapted model; for each criterion (Population, Intervention, Outcome, Study Approach), we compute contrastive similarity (inclusion–exclusion cosine) and a vagueness margin, which a Mamdani fuzzy controller maps into graded inclusion degrees with dynamic thresholds in a multi-label classification setting. A large language model (LLM) judge adjudicates highlighted spans with tertiary labels, confidence scores, and criterion-referenced rationales; when evidence is insufficient, fuzzy membership is attenuated rather than hard-excluded. In a pilot on an all-positive gold set ( full texts; chunks), the fuzzy system achieved document-level recall of 81.25% (Population), 87.50% (Intervention), 87.50% (Outcome), and 75.00% (Study Approach) with 95% Wilson confidence intervals, exceeding statistical baselines (recall 56.25–75.00%) and crisp baselines (recall 43.75–81.25%). Strict “All Criteria” inclusion was reached for 50.00% of articles, compared to 25.00% and 12.50% under the baselines. Cross-model agreement on justifications was 98.27%, and human–machine agreement was 96.07%. A pilot review showed 91% inter-rater agreement () with screening time reduced from ∼20 minutes to under 1 minute per article at a significantly lower cost. These results suggest that fuzzy logic, combined with contrastive highlighting and explainable LLM adjudication, delivers high recall, stable rationales, and end-to-end traceability. We emphasize that the reported recall figures were computed on an all-positive gold set (no excluded documents), so specificity, ROC/PR curves, and AUC are not defined here; the complete operating-characteristic analysis on a mixed-label corpus will be reported in a forthcoming study.Keywords: Systematic review automation, Full-text screening, Fuzzy logic (Mamdani), Multi-label text classification, Explainable AI, Large Language Models (LLMs), Contrastive semantic similarity, Biomedical NLP, Evidence synthesis Introduction Systematic reviews (SRs) are the cornerstone of evidence-based medicine and biomedical research, providing the most reliable synthesis of scientific knowledge for clinical, policy, and research decisions (Tawfik et al. 2019). Their rigor rests on pre-specified questions, comprehensive searches, explicit inclusion/exclusion criteria, and transparent appraisal. However, the scale of contemporary publishing has overwhelmed traditional workflows: recent estimates indicate that more than 75 clinical trials and 11 SRs appear each day (Bastian et al. 2010), and the pool of potentially eligible studies has expanded far beyond the available reviewer capacity (Qureshi 2023; Ofori-Boateng et al. 2024). In practice, teams now screen thousands of citations and frequently advance hundreds to full text (O’Mara-Eves et al. 2015; Marshall et al. 2018). Full-text screening as the bottleneck: Unlike title/abstract triage, full-text assessment requires careful reading of long, heterogeneous documents. Decisive information may be scattered across sections, buried in tables/figures, or described in discipline-specific or evolving terminology (Hasny et al. 2023; Gates et al. 2019). The task is laborious and error-prone. Reviewer fatigue and cognitive overload are common, and even with shared protocols, experts often disagree on eligibility at this stage (Gates et al. 2019). This disagreement reflects not only human inconsistency, but also a property of the data: full-text evidence is inherently ambiguous, context-dependent, and, in a precise sense, fuzzy. The inherently fuzzy nature of full-text decision-making By “fuzzy”, we mean that the signals that support inclusion/exclusion are graded, partial, and distributed. Populations are described broadly in one section and narrowly in another; interventions may be referenced indirectly; outcomes are reported with non-standard or shifting terminology. Criteria are seldom satisfied in an all-or-nothing manner; rather, eligibility emerges from partial matches and multi-dimensional semantic relationships. Consequently, two diligent reviewers can reasonably reach different conclusions on the same article, with consensus achieved only after a discussion that weighs context and degree—a phenomenon well documented for full-text screening (Gates et al. 2019). Limits of traditional and probabilistic approaches Crisp rule-based systems are explicit and transparent but poorly matched to graded contextual evidence; they cannot represent partial relevance or uncertainty (Esri 2025; Gates et al. 2019). Probabilistic machine learning offers improvements, but still reduces decisions to thresholds based on opaque scores and often lacks persuasive case-specific explanations (O’Mara-Eves et al. 2015; Marshall et al. 2018; Hasny et al. 2023). Recent advances in large language models (LLMs) improve coverage and flexibility. Yet LLMs alone do not guarantee consistent recall at SR standards, nor do they natively provide calibrated, graded decisions for the “gray-zone” cases that dominate full-text assessment (Guo et al. 2023; Kohl and Schön 2023). Why fuzzy logic is needed Fuzzy logic (Zadeh 1965) offers a human-aligned, principled calculus to represent degrees of membership and partial truth (Simkute et al. 2024; Wibowo and Deng 2012; Esri 2025; Bahadir Gökmen et al. 2025). In the context of full-text screening, much of the challenge arises not from simple ambiguity, which implies multiple possible interpretations, but from vagueness, where key eligibility signals are imprecise, graded, or previously non-measurable (Liu et al. 2024). Fuzzy logic rectifies this vagueness by transforming qualitative, diffuse evidence into quantifiable degrees of support, enabling systematic comparison across heterogeneous text spans (Liu et al. 2024). It encodes the interaction of signal strength and contextual nuance in a transparent rule base (Drnovšek et al. 2025), mirroring how expert panels reason: weighing multiple, partially supportive statements until a consensus degree of inclusion is reached. Crucially, fuzzy systems are both interpretable and auditable, allowing users to inspect membership functions, rule activations, and the exact path from text to decision (Mitrakas et al. 2025; Nzoundja Fapi et al. 2025). Innovations in fuzzy, explainable, and LLM-augmented screening We propose a full-text screening pipeline designed from first principles for fuzziness and explainability: Contrastive semantic highlighting: Articles are segmented and embedded; for each criterion, we compute inclusion-versus-exclusion similarity and highlight candidate spans, making the evidence traceable to concrete text. Contrastive semantic highlighting: Articles are segmented and embedded; for each criterion, we compute inclusion-versus-exclusion similarity and highlight candidate spans, making the evidence traceable to concrete text. Mamdani fuzzy inference: A Mamdani controller aggregates similarity strength and an explicit vagueness margin to produce graded inclusion degrees and support multi-label assignment (Bahadir Gökmen et al. 2025). Mamdani fuzzy inference: A Mamdani controller aggregates similarity strength and an explicit vagueness margin to produce graded inclusion degrees and support multi-label assignment (Bahadir Gökmen et al. 2025). LLM-as-judge: A domain-adapted LLM operates as a constrained validator over the fuzzy-selected spans rather than as an autonomous final arbiter. It reviews only pre-highlighted chunks, assigns a tertiary judgment (YES/NO/MAYBE) with a confidence score, and provides a short, criterion-referenced explanation. Its role is to refine and document the fuzzy decision, not to override protocol-based logic, and all document-level inclusion is evaluated against expert adjudication (Gu et al. 2025; Guo et al. 2023; Kohl and Schön 2023; Bedi et al. 2025; Dhungel 2025). LLM-as-judge: A domain-adapted LLM operates as a constrained validator over the fuzzy-selected spans rather than as an autonomous final arbiter. It reviews only pre-highlighted chunks, assigns a tertiary judgment (YES/NO/MAYBE) with a confidence score, and provides a short, criterion-referenced explanation. Its role is to refine and document the fuzzy decision, not to override protocol-based logic, and all document-level inclusion is evaluated against expert adjudication (Gu et al. 2025; Guo et al. 2023; Kohl and Schön 2023; Bedi et al. 2025; Dhungel 2025). Handling lack of evidence: When highlighted spans are judged insufficient at a later stage, their contribution is attenuated rather than forcing a brittle exclude/retain decision, preserving recall while demoting weak evidence. Handling lack of evidence: When highlighted spans are judged insufficient at a later stage, their contribution is attenuated rather than forcing a brittle exclude/retain decision, preserving recall while demoting weak evidence. This architecture targets high recall with bounded reviewer effort and an auditable chain from document text to decision. In this sense, “LLM-as-judge” is best understood as a structured validator and rationale generator that sits on top of the fuzzy controller, not as a free-standing decision-maker. The fuzzy layer determines which spans are evidentiary and how strongly they support each criterion; the LLM is then asked to confirm or contest those spans and to articulate why, under a fixed schema and prompt. This design mitigates hallucination risk, supports reproducibility, and keeps final inclusion aligned with pre-specified SR standards. Related work and comparative approaches Automated study selection has evolved along several methodological lines, including manual screening platforms, probabilistic machine learning, transformer-based classifiers, fuzzy logic systems, and, more recently, LLM-based prompting pipelines. Classic screening tools such as Covidence and Rayyan (Gates et al. 2019) provide structured workflows, conflict resolution, and tagging interfaces, but rely entirely on human judgment. They offer transparency but no automation beyond keyword filters. Active learning frameworks such as Abstrackr and ASReview (O’Mara-Eves et al. 2015; Marshall et al. 2018) were among the first to introduce machine learning into SR workflows, typically using SVMs or logistic models to reduce title/abstract workload. While effective at early-phase triage, these systems do not operate at full text and provide limited interpretability. A summary of representative systems and their strengths and limitations is provided in Table 1. Table 1Comparison of approaches for systematic review screeningApproachExamplesStrengthsLimitationsCrisp manual/rule-basedCovidence, Rayyan, keyword filters (Gates et al. 2019)High precision, explicit criteria, transparent; widely used in practiceLabor-intensive, inconsistent on ambiguous cases, no partial relevance, “maybe” still needs resolutionProbabilistic MLAbstrackr (SVM), BERT-based classifiers (O’Mara-Eves et al. 2015; Hasny et al. 2023; Marshall et al. 2018)Reduces workload, adapts to synonyms/implicit mentions, fast batch processing, active learning possibleBlack-box predictions, needs many labeled examples, output still thresholded to binary, lacks explanationFuzzy logic-basedFAME: Fuzzy Additive Models for Explainable AI (Bahadir Gökmen et al. 2025)Handles gradations/uncertainty, explainable, robust to heterogeneous reporting, aligns with expert reasoningRequires knowledge engineering, no large-scale precedent in SR, tuning needed, typically hybrid with MLLLM-basedGPT-4, recent prompt-based screening (Guo et al. 2023; Kohl and Schön 2023; Scherbakov et al. 2025; Delgado-Chaves et al. 2025; Teichert and Shah 2026)High comprehension, flexible, can explain in natural language, promising accuracy (in early studies)Recall not sufficient for SR needs, output inconsistent, explanation plausibility not guaranteed, scalability and cost issues Comparison of approaches for systematic review screening Transformer-based models, including BERT and domain-adapted variants (BioBERT, SciBERT), now represent state-of-the-art performance for automated SR screening (Hasny et al. 2023). These models often outperform classical probabilistic approaches and can be integrated into active learning loops. However, they express decisions through opaque logits and require substantial labeled training data. Their binary thresholding also struggles with borderline cases, and they provide no structured justification tied to protocol criteria. Recent efforts add attention visualization or saliency maps, but these explanations remain coarse and difficult to relate directly to eligibility criteria. Parallel to probabilistic and transformer approaches, fuzzy logic has been explored for text classification tasks that exhibit graded or ambiguous signals. Recent work demonstrates the effectiveness of fuzzy neural networks and hierarchical fuzzy classifiers in multi-label and noisy-text settings (Ahmad et al. 2025; Drnovšek et al. 2025). Fuzzy additive models for explainable AI (Bahadir Gökmen et al. 2025) and other rule-based fuzzy systems offer transparent reasoning pathways, enabling partial membership and interpretable aggregation of uncertain evidence. A recent comprehensive survey of interpretable fuzzy methods (Zhang et al. 2026) further highlights the central role of fuzzy rule systems, fuzzy sets, and neuro-fuzzy architectures in transparent decision-making, and documents key trends such as accuracy–interpretability trade-offs and human-in-the-loop reasoning. This literature reinforces the suitability of fuzzy inference as a principled foundation for explainable screening in settings characterized by ambiguity and graded evidence. However, existing applications focus primarily on short text or structured attributes rather than long-form biomedical literature, and they rarely integrate with semantic embedding models or LLMs. LLM-based screening pipelines have gained interest due to models’ ability to parse complex scientific language. Early prompting frameworks show promising recall for title/abstract screening (Guo et al. 2023; Kohl and Schön 2023). Some recent studies apply LLMs for broader document triage, classification, or scenario interpretation in adjacent fields (Teichert and Shah 2026). However, LLMs often exhibit inconsistent judgments, lack grounding in protocol definitions, and may hallucinate rationale text. Moreover, LLM outputs are typically non-deterministic and difficult to audit, limiting their direct use as stand-alone decision-makers for SR workflows. Hybrid systems combining rule-based logic with ML or LLM components have emerged as another promising direction. These systems attempt to balance transparency with flexibility, using deterministic rules for constraint enforcement and ML models for semantic inference. Such designs mirror expert committee reasoning in SRs, but practical deployments remain limited. A notable gap in prior work is the integration of semantic highlighting, fuzzy aggregation, and LLM adjudication within a single, auditable pipeline. Our contribution builds on and extends these strands. We combine domain-adapted embeddings for contrastive highlighting, a Mamdani fuzzy controller for aggregating graded signals, and a constrained LLM-as-judge for validating and explaining only the fuzzy-selected spans. Unlike transformer classifiers, our method does not require large labeled datasets, and unlike pure LLM pipelines, it constrains the model within a structured, auditable decision process. The design targets high recall, transparent rationale generation, and the ability to trace each inclusion decision to specific text spans. These capabilities are not jointly offered in prior work across SR automation platforms, transformer-based approaches, fuzzy logic methods, or LLM-based systems. Evaluation overview We evaluate the proposed system using a multi-metric battery appropriate for screening tasks: recall (on an all-positive gold set), reviewer agreement, explanation quality, and efficiency; and, for future mixed-label benchmarks, specificity and threshold-free discrimination (ROC/PR curves and AUC), alongside calibration and ablations. This combination aligns with best practices in explainable AI and targets real-world deployability. In summary, the exponential growth of the literature and the fundamentally graded nature of full-text evidence argue for a shift from crisp or purely probabilistic screening to a fuzzy, explainable paradigm. By integrating contrastive highlighting, Mamdani inference, and LLM adjudication, our approach seeks to deliver high recall, transparent rationales, and scalable operations suitable for living SRs. This study is a methodological proof-of-concept that introduces and evaluates an automated full-text screening pipeline for systematic reviews; it is not itself a systematic review or tutorial. Methodology Full-text semantic highlighting, fuzzy logic screening, and explainable LLM judgment We build a pipeline that operationalizes the intrinsically fuzzy, context-dependent nature of full-text eligibility. The system combines: (i) semantic highlighting to surface evidence spans tied to explicit criteria, (ii) a Mamdani fuzzy inference layer to map similarity and ambiguity into graded inclusion scores, and (iii) LLM-based adjudication to provide auditable, criterion-referenced rationales. The design targets high recall with bounded reviewer effort and end-to-end traceability: every automated decision is linked to concrete text and a short explanation. The overall workflow is depicted in Figure 1. Overview of the automated full-text screening pipeline. The process consists of four stages: (1) text chunking and embedding, (2) semantic similarity scoring and highlighting, (3) fuzzy logic selection, and (4) LLM-based explainable validation and adjudication. User-submitted documents are divided into chunks, semantically compared to inclusion/exclusion criteria, filtered by a Mamdani fuzzy inference system, and finally validated with LLM-based explanations and labels Text chunking, embedding, and semantic highlighting Parsing and chunking: Each PDF is parsed into a linearized text stream while preserving section boundaries when available (e.g., section titles, headings, captions) (Wang et al. 2025). We segment into overlapping “chunks” of 3–5 sentences with a stride of 1–2 sentences, which balances local coherence with coverage of cross-sentence evidence (Rodriguez et al. 2025). Let denote the documents and the chunks of . Embedding: Each chunk c is mapped to a dense vector using a domain-adapted embedding model (e.g., OpenAI text-embedding-3-large (Wang et al. 2024; Petukhova et al. 2025)): where is the embedding function. Criteria texts (verbatim inclusion and exclusion definitions) are embedded once and reused, enabling efficient, batched similarity. Highlighting. Given criterion-specific vectors, candidate chunks are highlighted by semantic proximity (defined below). Highlights are retained with offsets to support later visualization and auditing. This step surfaces where the model found evidence, not just whether it exists. Contrastive semantic similarity scoring Contrastive score: For each criterion k, we embed the inclusion and exclusion specifications, and . The relevance of chunk c to k is the subtracted cosine similarity: which rewards alignment with inclusion while penalizing overlap with exclusion. This contrastive formulation reduces false positives from generic or off-target matches and stabilizes ranking across heterogeneous writing styles. Normalization and stability: Cosine values are bounded in [-1,1]; the subtraction yields . In practice, we clip to a plausible band for scientific text (e.g., [-0.10,0.15]) to avoid downstream rule inactivity and to keep the fuzzy universes compact (cf. Figure 3). Triangular membership functions used in the Mamdani fuzzy inference system. (Left) Score membership functions for subtracted cosine similarity. (Center) Margin membership functions. (Right) Decision membership functions mapping to reject/maybe/accept; the vertical dashed line indicates the default cutoff Margin calculation and ambiguity quantification To quantify local ambiguity, we compute Here denotes the ambiguity margin used later to support multi-label assignment. A small positive indicates that multiple criteria explain the chunk with similar strength (for example, text describing both population and intervention). A zero or negative signals an unstable or tied top score, typically arising from noise or weak evidence. Both and are passed to the fuzzy layer. Fuzzy logic (Mamdani) selection A Mamdani fuzzy inference system maps to a graded inclusion degree for each plausible criterion k (Lilly 2010; Shirgir and Farahmand-Tabar 2025; Salameh et al. 2025) (Figure 2). For multi-label cases (small ), the fuzzy controller is applied independently to each plausible criterion k, using in place of . Membership functions for inputs and outputs are visualized in Figure 3. Block diagram of a Mamdani fuzzy inference system. Crisp inputs are fuzzified, evaluated by an inference engine and rule base, and defuzzified to a scalar score Fuzzification: Inputs and are mapped to linguistic terms (low/medium/high for score; small/medium/large for margin) via standard triangular membership functions: for parameters a