This paper presents a scalable and auditable pipeline based on fuzzy logic to address expert screening, a major bottleneck in systematic reviews (SRs). Instead of using traditional dichotomous inclusion/exclusion criteria, inclusion/exclusion is reframed as a fuzzy decision problem, dividing the literature into overlapping chunks and embedding them with a domain-adaptive model. For each criterion (Population, Intervention, Outcome, Study Approach), contrastive similarity (inclusion-exclusion cosine) and ambiguity margin are computed, and a Mamdani fuzzy controller is used to map these to graded inclusion scores with dynamic thresholds in a multi-label setting. A large-scale language model (LLM) assesses highlighted intervals with tertiary labels, confidence scores, and criterion-referenced evidence, reducing fuzzy membership instead of excluding them when evidence is insufficient. In a pilot study targeting the POPCORN (Population Health Modeling Consensus Reporting Network for Noncommunicable Diseases) positive gold set (16 full-text articles; 3,208 chunks), the proposed fuzzy system achieved high recall (Population 81.3%, Intervention 87.5%, Outcome 87.5%, Study Approach 75.0%), surpassing the statistically significant baseline. The inclusion rate for papers meeting all criteria was 50.0%, higher than the baselines (25.0% and 12.5%). The screening time was reduced from approximately 20 minutes to less than 1 minute, significantly reducing costs. In conclusion, the system combining fuzzy logic, contrastive highlighting, and LLM judgment provides high recall, robust evidence, and end-to-end traceability.
