Health care is strongly influenced by uncertainty. Clinical and public health decisions are routinely made under conditions of incomplete, ambiguous, or imprecise information, arising not only from individual patient variability but also from the complexity of health care systems and the processes through which real-world data are generated [1]. Such uncertainty encompasses both stochastic variability and epistemic constraints and is further amplified by heterogeneous, noisy, and nonlinear data from electronic health records, diagnostic imaging, physiological signals, and population-level monitoring systems [2]. Under these conditions, conventional statistical approaches—typically relying on fixed thresholds, linearity, and prespecified functional forms—often struggle to represent gradual transitions, ambiguous diagnostic boundaries, and context-dependent relationships that characterize real-world clinical data [1-3].

Causal inference emerged as a methodological approach for estimating the effects of exposures or interventions on health outcomes, explicitly addressing the limitations of purely associational analyses [4-10]. Rather than focusing on prediction, this approach centers on counterfactual questions—what would be expected to occur under hypothetical alterations in treatment or exposure—by making causal assumptions explicit and, in principle, empirically assessable [9,11-13]. This perspective is particularly relevant in health care and public health, where randomized controlled trials are frequently impractical and observational data constitute the primary source of evidence [9,11,13,14]. In such settings, causal reasoning is commonly formalized using directed acyclic graphs, which encode assumptions about causal structure, confounding, and intervention pathways, thereby enabling principled identification of causal effects [15-17].

Recent developments have emphasized the central role of explicit study design in strengthening causal inference from observational data. Target trial emulation (TTE) clarifies the causal question by prespecifying the key protocol components of a hypothetical randomized trial—including eligibility criteria, treatment strategies, time zero, follow-up, and outcomes—prior to analysis, thereby aligning observational studies with the core principles of randomized experiments [17-20]. While this design-oriented framework can reduce avoidable biases and enhance interpretability, it does not by itself ensure valid effect estimation. In practice, TTE still requires appropriate identification assumptions and estimation strategies and may remain vulnerable to challenges such as model misspecification or limited flexibility when representing complex data structures [17].

Despite advances in causal inference and design-oriented approaches, substantial challenges persist at the estimation stage when analyzing complex health care data. Even when causal questions are explicitly defined, commonly used estimation methods often rely on inflexible functional assumptions, sharply delineated variables, and correctly specified models—conditions that are difficult to sustain in high-dimensional and heterogeneous clinical environments [3,8,11]. As a result, a methodological gap remains between rigorously specified causal designs and the flexible representation of nonlinearity, gradual clinical thresholds, and uncertainty inherent in observational health data, limiting the applicability of traditional causal models in complex real-world settings [1,2,11,14,21].

As a response to the demand for flexible representations of uncertainty and nonlinearity, fuzzy logic has been adopted as a modeling paradigm in health care research, grounded in earlier theoretical developments on vagueness and graduality. Central to this evolution was Zadeh’s introduction of fuzzy sets as a generalization of classical set theory, in which membership is defined by degrees rather than binary inclusion [22-24]. This formulation provided a formal mathematical basis for representing ambiguity, partial truth, and gradual transitions in complex systems, thereby enabling the representation of phenomena that cannot be adequately captured using crisp categories. Building on this foundation, subsequent developments extended fuzzy sets into operational fuzzy logic systems, particularly through rule-based inference mechanisms that support reasoning with linguistic variables and imprecise conditions [25].

In clinical contexts, this representational flexibility facilitates the translation of gradual and linguistically defined clinical concepts into implementable computational models. Building on these foundations, a range of fuzzy logic–based approaches—including fuzzy inference systems (FIS), adaptive neuro-fuzzy inference systems (ANFIS), fuzzy cognitive maps (FCM), Takagi-Sugeno models, and fuzzy clustering—have been applied across diverse health care domains. These applications span infectious diseases [26-34], cardiology [35-42], oncology [43-49], obstetrics [50-52], mental health [53,54], and occupational health and safety [55-59].

More recently, fuzzy logic–based approaches have increasingly been combined with machine learning and artificial intelligence techniques to enhance predictive performance, scalability, and automation in health care applications [60,61]. Despite this growing convergence, the extent to which such hybrid models explicitly engage with causal reasoning—through the definition of counterfactual estimands, formal identification strategies, and transparent causal assumptions—remains inconsistently reported in the literature. Against this background, this systematic review aimed to evaluate and synthesize evidence on the application of fuzzy logic–based approaches for causal inference in health care.

During the systematic search conducted between March and June 2025, a total of 621 records were identified across 6 electronic databases. After removal of duplicates and screening of titles, abstracts, and full texts, 37 studies published between 2014 and 2025 met the inclusion criteria and were retained for final synthesis. The PRISMA 2020 flow diagram (Figure 1) details the study selection process, including the number of records screened, excluded, and included at each stage of the review.

Table 2 lists fuzzy logic–based methodologies from 37 studies, showing the literature’s methodological diversity. The table defines each approach, its main analytical role, common application in reviewed studies, and frequency of use.

Table 3 summarizes the 37 studies included in the final synthesis, spanning health care domains such as infectious diseases, cardiovascular conditions, cancer, mental health, occupational health, and preterm birth. Across studies, the most frequently applied fuzzy approaches were FIS (Mamdani type), ANFIS, fuzzy analytic hierarchy process (FAHP) variants, and FCM. Sample sizes varied widely, ranging from fewer than 100 participants to large-scale public datasets exceeding 1000 cases, with data sources including institutional or hospital records, expert-based judgments, and simulated data.

Fourteen studies used direct comparator methods, most commonly logistic regression, decision trees, or ensemble classifiers, whereas the remaining studies relied on baseline comparisons or did not include external benchmarks. Performance was typically reported using accuracy or AUC, with sensitivity and specificity included in selected cases. Importantly, only a minority of studies explicitly framed their analyses within formal causal inference paradigms, while most remained primarily predictive or associative in scope.

The temporal distribution of the included studies shows an uneven pattern over the past decade, with episodic increases rather than a steady growth trajectory, culminating in a pronounced peak in 2024 (Figure 2).

The most frequently addressed conditions were infectious diseases (n=10) [26-34,70], cardiovascular diseases (n=7) [35,36,38-42], cancer (n=7) [43-49], occupational health and safety (n=5) [55-59], mental health (n=2) [53,54], and preterm birth (n=2) [50,51]. Additional studies fell into miscellaneous categories [66,68,69].

Regarding data sources, most studies used institutional or hospital datasets (n=18) [31,35,36,41,44-48,50,51,55,56,66-71], while 9 relied on public datasets [28-30,32,33,38,40,49,57], 5 reported expert-based data [27,34,39,54,59], 3 reported mixed sources [26,53,58], and 2 used simulated or synthetic data [42,43].

Sample sizes varied considerably across studies: 13 used large datasets (n≥1000) [26,28-32,36,38,45,46,51,67,68], another 10 relied on medium-sized samples (n=100 to 999) [35,40,41,44,49,50,53,58,66,69], 10 used small datasets (n<100) [34,43,47,48,54-57,59,68,70], and in 4 studies, the sample size was not applicable [27,33,39,42].

A broad array of fuzzy logic techniques was identified across the included studies, reflecting substantial methodological heterogeneity. The most commonly used methods were FIS and their variations (n=8) [26,29,33,35,55,56,58,70], frequently implemented using Mamdani-type structures; followed by the FAHP (n=6) [26,31,32,39,50,67], FCM (n=5) [28,43,44,54,66,69], adaptive neuro-fuzzy systems (n=3) [41,45,48], typically combining neural architectures with fuzzy rule bases for improved learning capacity and hybrid fuzzy approach combined with multicriteria decision models (n=3) [27,38,57].

Other used models were fuzzy clustering (C or K means) (n=2) [51,67], fuzzy-trace theory (n=2) [46,49], fuzzy-set qualitative comparative analysis (fsQCA) (n=2) [30,53], Takagi-Sugeno models (n=1) [66], fuzzy failure mode and effects analysis (n=1) [59], mediative fuzzy logic (n=1) [42], fuzzy evidential reasoning (n=1) [34], likelihood-fuzzy analysis (n=1) [47], and profile-based fuzzy association rule mining (n=1) [40].

Among the studies reviewed, 14 conducted direct comparative evaluations against traditional methods such as logistic regression, decision trees, or standard statistical models [28,39-41,43-45,47,48,57,59,66,68,69] and 6 studies used baseline comparisons, typically involving simple pre/post assessments without an external benchmark [26,31,32,46,50,56]. In contrast, 17 studies applied fuzzy modeling in isolation, without any form of benchmarking or comparator method, relying solely on internal outputs to assess performance [27,29,30,33-36,38,42,49,51,53-55,58,67,70]

Between the studies that conducted direct comparative evaluations, 5 reported that fuzzy logic models outperformed traditional methods, including statistical classifiers and machine learning algorithms. These included Mahmoodi et al [44], who achieved 95.8% accuracy in gastric cancer prediction using FCM; Yılmaz et al [45], who obtained 94.6% accuracy with a neuro-fuzzy model for lung cancer; Subramanian et al [43], who reported 94.3% overall accuracy using a layered FCM for breast cancer risk; Sabahi [39], who introduced a bimodal FAHP model with accuracies above 85%; and Saleh et al [41], whose ANFIS classifier outperformed other ensemble models in diabetic retinopathy detection.

Three studies showed that fuzzy models yielded comparable or slightly superior performance relative to conventional methods. Argyropoulos et al [68] reported equivalent AUC values for both fuzzy logic and logistic regression models in predicting acute kidney injury, while Pota et al [47] found similar predictive accuracy between likelihood-fuzzy analysis and naïve Bayes classifiers in radiotherapy toxicity. Stanković and Stanković [48] also demonstrated that a neuro-fuzzy system marginally outperformed an artificial neural network in predicting prostate cancer survival.

The remaining 6 studies—Amirkhani et al [66], Yavari et al [40], Mohandes et al [57], Benito et al [28], Sümbül-Şekerci et al [69], and Demir and Sabır [59]—involved direct comparisons but did not report sufficient methodological or statistical detail to clearly assess the relative effectiveness of the fuzzy approach. To visually summarize the comparative performance of fuzzy logic models versus conventional statistical approaches, Figure 3 presents reported accuracy values from studies that provided quantifiable metrics. Only those studies with explicit accuracy comparisons were included, enabling a focused assessment of relative predictive performance across diverse health care domains.

Across these studies, common performance metrics included accuracy (84%‐95.8%), AUC (0.70‐0.95), and error measures such as root mean square error, mean absolute error, and mean squared error. These results underscore the adaptability of fuzzy modeling to clinical decision-making contexts marked by uncertainty, incomplete data, and the need for interpretability.

In terms of causal inference, conceptual approaches varied across the studies. While most of the studies addressed high-complexity settings involving multiple interacting variables, only two explicitly adopted formal causal inference frameworks. These included Lee et al [30], who used fsQCA with sufficiency and necessity thresholds; and Mohandes et al [57], who implemented a hybrid interval-valued intuitionistic fuzzy DEMATEL-ANP (decision-making trial and evaluation laboratory analytic network process) model with cross-validation.

Six additional studies [28,38,44,53,54,66] simulated causal mechanisms using methods such as iterative expert-based system mapping or FCM. However, none of these studies explicitly operationalized a formal causal inference framework grounded in counterfactual reasoning or directed acyclic graphs. Instead, causal assumptions were inferred through expert consensus or embedded in the structure of fuzzy systems.

The remaining 29 studies used fuzzy logic primarily for predictive or associative analysis [26,27,29,31-36,39-43,45-51,55,56,58,59,67-70], with causal relationships often left implicit, untested, or loosely derived from domain-specific knowledge alone.

Table 3 provides a detailed synthesis of the 14 studies that directly compare fuzzy logic models with traditional statistical or machine learning methods. Most of these studies reported performance gains for fuzzy approaches, particularly in cancer [43-45] and cardiovascular domains [39,41]. In several cases, fuzzy models offered not only higher accuracy or sensitivity but also enhanced interpretability. Others showed broadly comparable results with added value in robustness [47,48,68]. A smaller group reported either mixed outcomes or limited statistical detail, emphasizing interpretability and methodological novelty over raw predictive gains [28,40,57,59,66,69].

Collectively, the evidence summarized in Table 4 indicates that fuzzy logic–based approaches have been evaluated against conventional methods in a limited subset of studies, yielding heterogeneous results and variable reporting quality. Although several comparative assessments suggest potential advantages in managing uncertainty and enhancing interpretability, the absence of systematic benchmarking and the predominance of predictive objectives preclude definitive conclusions regarding comparative effectiveness.

While the findings indicate that fuzzy logic–based approaches are frequently applied in predictive health care modeling, the strength of the available evidence must be interpreted considering methodological quality and risk of bias. Seventeen studies were evaluated using the PROBAST+AI tool, specifically designed for assessing bias in prediction model studies [65]. Of these, 9 were rated as having a high risk of bias [27,34,35,39,43,44,47,66,67], and 8 studies were rated as moderate risk [28,33,40,41,45,48,68,69], standing out for more robust validation procedures, detailed variable handling, and partial transparency. None achieved a low-risk rating.

Of the 20 studies assessed using the JBI checklist [64] for analytical cross-sectional designs, 5 were rated as having low risk of bias [29,30,46,49,53], while the remaining 15 [26,31,32,36,38,42,50,51,54-59,68,70] were classified as moderate risk (Figure 4).

To examine the distribution of fuzzy logic techniques across health care applications, a cross-tabulated synthesis was conducted. As shown in Figure 5, the most frequently applied approaches were FIS, FCM, and ANFIS, followed by FAHP, fsQCA, fuzzy evidential reasoning, and Takagi-Sugeno models. The use of these techniques varied across application domains, with oncology, infectious diseases, cardiovascular health, and mental health exhibiting the highest methodological diversity.

The nature of causal engagement across the included studies spanned a continuum from explicitly formalized causal frameworks to approaches in which causal reasoning remained implicit or embedded within expert-driven or structurally defined fuzzy models. Only two studies (2/37, 5.4%) explicitly addressed causal questions using formal causal inference methodologies. A small subset relied on inferred causal structures derived from expert knowledge or fuzzy cognitive maps (6/37, 16.2%). In contrast, most studies primarily implemented predictive or associative modeling approaches, where causal interpretation was not formally specified and was instead inferred indirectly from model structure, expert judgment, or post hoc interpretation (29/37, 78.4%). This distribution highlights substantial heterogeneity in how causal principles are operationalized across fuzzy logic–based applications in health care.

This systematic review synthesized evidence from 37 studies published between 2014 and 2025 that used fuzzy logic–based methodologies in health care settings with explicit or implicit causal objectives. The included studies span a wide range of clinical and public health domains, including infectious diseases, cancer, cardiovascular diseases, occupational health and safety, mental health, and preterm birth, underscoring the broad applicability of fuzzy modeling to diverse health-related problems. Across domains, the most frequently reported approaches were FIS, ANFIS, FAHP, and FCM. Rather than indicating methodological convergence, this distribution reflects context-dependent adaptations of fuzzy logic to address uncertainty, nonlinearity, and expert-guided reasoning in complex health care environments.

Only a limited subset of studies conducted direct comparative evaluations between fuzzy logic–based models and conventional statistical or machine learning approaches. Among the 14 studies that included explicit comparators, 5 reported superior performance of fuzzy models [39,41,43-45]—most frequently in cancer and cardiovascular applications—while 3 demonstrated broadly comparable results [47,48,68]. The remaining 6 studies provided comparative analyses with insufficient methodological or statistical detail to support firm conclusions regarding relative effectiveness [28,40,57,59,66,67,69]. Importantly, most included studies relied on internal validation procedures, baseline comparisons, or expert-defined structures without external benchmarks, often using small- to medium-sized datasets. This pattern limits the generalizability of reported performance gains and indicates that, while fuzzy approaches may perform competitively in specific contexts characterized by nonlinearity or uncertainty, evidence supporting consistent superiority over conventional methods remains limited and heterogeneous.

Causal inference was explicitly operationalized in only a small proportion of the included studies. Specifically, two investigations adopted formal causal inference frameworks: Lee et al [30] used fsQCA, explicitly modeling configurations of necessary and sufficient conditions at the population level. Mohandes et al [57] implemented a hybrid interval-valued intuitionistic fuzzy DEMATEL-ANP approach to structurally identify and prioritize causal drivers in occupational safety systems. In both cases, causal claims were grounded in transparent methodological procedures, explicit thresholds, and internally coherent validation strategies, rather than inferred post hoc from predictive performance.

Beyond these two studies, causal reasoning was indirect. Six additional investigations relied on expert-based mappings, FCM, or influence structures to simulate causal mechanisms without formally testing necessity, sufficiency, or counterfactual dependence [28,38,44,53,54,66,69]. In most studies, fuzzy logic was applied primarily for predictive or associative purposes, with causal assumptions embedded implicitly within model architecture or domain expertise rather than explicitly articulated or empirically evaluated.

Taken together, these findings reveal a marked disconnect between the theoretical capacity of fuzzy logic to represent causal structure and its prevailing empirical use in health care research. This gap appears to reflect not inherent conceptual limitations of fuzzy methods, but rather broader issues related to study design, validation practices, and reporting rigor, which constrain the translation of fuzzy modeling from predictive decision support to explicit causal inference.

Risk of bias constituted a major limiting factor across the included studies. Among those evaluated with PROBAST+AI [65], none achieved a low-risk rating, with most classified as moderate or high risk, while only a small proportion of studies assessed using the JBI checklist [64] were rated as low risk.

In contrast to much of the existing literature, which has primarily emphasized predictive accuracy or isolated clinical applications, the present review integrates formal risk-of-bias assessment with thematic synthesis to jointly evaluate reported performance, methodological rigor, and the explicitness of causal assumptions. This perspective highlights both the strengths and current limitations of fuzzy logic–based approaches: while they provide interpretable, rule-based models well suited to ambiguity and nonlinearity, their application within explicitly causal analytical frameworks remains limited and inconsistent.

These conclusions must be interpreted considering several important limitations, including substantial heterogeneity across health care domains, modeling strategies, and outcome measures, which precluded quantitative meta-analysis; inconsistent reporting practices, such as limited use of comparator models and incomplete outcome reporting; and the frequent reliance on small- to medium-sized datasets without external validation. Collectively, these factors reduce the overall certainty and generalizability of the current evidence base.

Despite these limitations, the findings carry important implications for both research and practice. Fuzzy systems appear particularly well suited to health care and policy contexts characterized by incomplete data, multidimensional interactions, and a strong demand for interpretability. Their capacity to encode expert knowledge and tolerate imprecision supports their use in applications such as risk stratification, early diagnosis, and context-sensitive prioritization. Realizing this potential, however, will require methodological consolidation, including greater standardization in reporting, more consistent use of comparator frameworks, and external validation across real-world datasets. Importantly, integration with formal causal frameworks—such as directed acyclic graphs or structural causal models—offers a pathway to strengthen causal interpretability while preserving the distinctive advantages of fuzzy reasoning.

In parallel, recent advances in artificial intelligence have largely emphasized the automation of data extraction, measurement, and pattern recognition in clinical settings, particularly through machine learning and computer vision–based applications [71-76]. While these approaches have improved efficiency and scalability, they remain predominantly oriented toward prediction rather than causal inference. Addressing this gap requires analytical frameworks that move beyond automation to explicitly represent causal structure, intervention contrasts, and temporal assumptions.

In this context, future research would benefit from explicitly incorporating TTE [17-20] when applying fuzzy logic to observational health care data. TTE provides a principled framework for specifying causal estimands, temporal ordering, and hypothetical interventions, thereby addressing key sources of bias that remain unresolved in many fuzzy-based applications. By defining eligibility criteria, treatment strategies, follow-up periods, and causal contrasts a priori, TTE can situate fuzzy, rule-based models within transparent causal designs—an approach that is particularly relevant in real-world health care settings where randomized trials are often infeasible.

Viewed in this way, fuzzy logic should not be considered merely an auxiliary modeling technique, but a potential component of hybrid causal approaches in health care. When interpretability and causal structure are integrated into model design rather than treated as secondary considerations, fuzzy systems may help bridge the gap between statistical prediction and meaningful causal explanation. Advancing this agenda will require further methodological refinement, interdisciplinary collaboration, and a move toward more coherent and explicitly causal research programs in complex health systems.