Researchers have introduced an explainable artificial intelligence framework designed to detect and eliminate hidden data inconsistencies in biomedical text classification, demonstrating substantial improvements in classification performance across oncological research literature. Published in BMC Artificial Intelligence, the study addresses a critical bottleneck in biomedical natural language processing, where conventional classification pipelines frequently stumble due to semantic misalignment, noisy document labels, and structural contradictions embedded in biomedical graphs. Manual error analysis has historically proven labor-intensive and prone to human oversight, restricting the practical utility of automated classifiers within clinical decision support systems. To overcome these limitations, the authors developed a dual-perspective framework that decouples unsupervised semantic clustering from supervised classification. By systematically isolating and removing discordant data points rather than relying on arbitrary data pruning, the optimized architecture elevated overall classification accuracy from forty-six percent to ninety-one percent, while boosting the corresponding F1-score from fifty percent to eighty-nine percent, validated through rigorous statistical testing.
The empirical investigation evaluated document categorization across three primary malignancy categories: colon cancer, lung cancer, and thyroid cancer. The primary corpus comprised 7,569 peer-reviewed publications curated for biomedical text classification, incorporating extended abstracts and research papers exceeding six pages, distributed among 2,579 colon cancer records, 2,180 lung cancer articles, and 2,810 thyroid cancer studies. To test system adaptability against contemporary academic literature, the investigators curated a secondary benchmark consisting of titles and abstracts from 900 cancer research manuscripts published during 2024 and 2025, balanced with 300 entries per malignancy class. Before model training, the raw text underwent a rigorous preprocessing pipeline to standardize linguistic representations. The text was parsed into discrete tokens, followed by stop-word removal to eliminate common words lacking diagnostic context. Crucially, the researchers deliberately excised explicit anatomical keywords directly tied to the target classes—specifically deleting instances of colon, thyroid, and lung cancer—concluding the preparation phase with lemmatization to reduce words to canonical dictionary forms.
To capture relational dependencies governing biomedical statements, the framework extracts grammatical entities to build structured knowledge graphs from unstructured prose. Using the Spacy natural language processing library and its dependency parser, the pipeline decomposes each sentence into grammatical triples consisting of subjects, verbs, and objects. In the resulting network topologies, extracted nominal entities functioning as subjects and objects serve as individual graph nodes, while connecting verbs describing biological interactions and medical actions form directed relational edges. This subject-verb-object parsing converts unstructured medical literature into interpretable semantic networks capable of exposing latent associations between pathological terms, therapeutic interventions, and anatomical sites. The investigators modeled and inspected these topological structures using specialized software packages including StellarGraph, NetworkX, and PyViz. By translating narrative sentences into graph representations, the methodology ensures biomedical concepts are modeled as interconnected nodes within a broader relational web, establishing the structural foundation required for downstream embedding and semantic verification.
Once knowledge graphs were constructed, the researchers applied the Node2Vec graph embedding algorithm to translate discrete graph nodes into continuous, low-dimensional feature vectors that preserve local and global neighborhood structures. The Node2Vec framework explores network spaces through a biased random walk procedure governed by hyperparameters that balance breadth-first exploration against depth-first traversal across graph vertices. By executing repeated random walks of defined length from each node, the algorithm samples neighborhood contexts and optimizes vector representations using stochastic gradient descent. This flexibility allows node2vec to organize nodes based on their network roles and communities they belong to, adapting to various network structures and enabling more accurate representations. The resulting twenty-dimensional embeddings encapsulate complex relational properties that surface latent discrepancies in data annotations. These learned node representations are subsequently supplied to a supervised Support Vector Machine classifier, which leverages kernel transformations to establish optimal separating hyperplanes across the high-dimensional feature landscape.
Operating in parallel with the graph embedding pipeline, the framework deploys BERTopic to generate unsupervised semantic clusters directly from textual documents, establishing an independent linguistic perspective against which graph-based classifications can be cross-examined. BERTopic transforms sentences into dense continuous vectors using pre-trained Sentence-BERT models optimized for semantic similarity. Because raw sentence embeddings reside in high-dimensional vector spaces where conventional distance metrics lose discriminatory power, the pipeline implements Uniform Manifold Approximation and Projection to reduce dimensionality while preserving both local and global geometric topologies. Following dimensionality reduction, the hierarchical density-based clustering algorithm HDBSCAN groups the compressed embeddings into distinct thematic clusters while segregating unclassifiable noise and peripheral outliers. In comparative architectural evaluations, the investigators explored two operational clustering configurations: an unconstrained setting permitting arbitrary topic numbers, and a constrained setting enforcing a three-cluster ceiling matching the target cancer categories. Experimental validation demonstrated that confining cluster generation to the three predefined diagnostic classes produced superior downstream classification alignment.
To reconcile unsupervised clusters with ground-truth clinical annotations, the researchers developed an automated mapping mechanism designed to assign each discovered semantic cluster to its corresponding cancer classification. While simple Class-to-Cluster Assignment relying on majority frequency counts was tested, the authors introduced a collaborative filtering framework termed Recommender-System Assignment implemented via the Python Surprise library. In this formulation, individual documents were modeled as item identifiers, BERTopic clusters functioned as user identifiers, and human ground-truth diagnostic labels served as rating values. The recommender model decomposed the sparse cluster-label association matrix into lower-dimensional latent factor matrices representing cluster features and label characteristics, incorporating quadratic regularization to prevent overfitting amid lexical sparsity. This matrix factorization approach uncovered latent relationships between clusters and labels with an initial prediction accuracy of eighty-six percent.
The centerpiece of the methodology is the Explainable Inconsistency Detection module, an algorithmic engine designed to identify, justify, and resolve points of divergence between semantic clustering and supervised classification. Inconsistency is formally defined as any scenario where the predicted class generated by the Support Vector Machine operating on node embeddings contradicts the aligned class assigned through BERTopic semantic clustering. When such divergences occur, the framework flags the affected data points as latent labeling errors, semantic contradictions, or noisy graph artifacts. To handle feature data assimilation, the authors evaluated two distinct information propagation strategies: projecting document-level topic labels down to constituent grammatical nodes, or aggregating node-level predictions upward to establish a single dominant label for an entire document graph. Finding that node-level analysis yielded superior precision, the authors adopted a targeted exclusion policy, excising conflicting nodes from the embedding dataset. Furthermore, they introduced a bounded removal parameter, beta, calibrated to zero-point-seven to selectively remove seventy percent of identified inconsistencies, successfully preserving dataset volume.
Systematic experimental evaluations conducted on both cancer corpora confirmed the decisive operational benefits of targeted inconsistency exclusion over traditional classification pipelines. Before inconsistency elimination, the baseline Support Vector Machine classifier achieved an accuracy of only forty-six percent and an F1-score of fifty percent when processing raw node embeddings. Following the removal of conflicting nodes via Recommender-System Assignment optimization, classification accuracy surged to ninety-one percent, accompanied by an F1-score of eighty-nine percent on the primary dataset. Class-specific performance evaluations using Receiver Operating Characteristic curves revealed substantial discriminative gains, with the Area Under the Curve for Class 0 climbing from zero-point-fifty-five to zero-point-ninety-four, while Class 2 expanded from zero-point-fifty-seven to zero-point-ninety-five. Class 1 exhibited a more modest Area Under the Curve improvement, rising from zero-point-fifty-three to zero-point-fifty-eight. To confirm that performance gains were not an artifact of reduced sample sizes, control trials removed equivalent data volumes randomly. Two-sample t-tests yielded p-values several orders of magnitude below zero-point-zero-five, confirming statistical significance.
Beyond quantitative performance metrics, the Explainable Inconsistency Detection framework provided critical interpretability by pinpointing specific topological and semantic anomalies responsible for model confusion. Graph centrality calculations revealed dramatic structural divergences between consistent and inconsistent lexical nodes across the constructed biomedical networks. Consistent terms demonstrated a mean degree centrality of one-point-three-eight-five times ten to the negative third power, whereas inconsistent terms exhibited a significantly lower degree centrality of two-point-zero-five times ten to the negative fifth power. Network visualizations demonstrated that ambiguous terms frequently established misleading structural bridges between disparate tumor types; for example, the term lung cancer was errantly linked to both PTEN expression and liver cancer, generating topological confusion. Similarly, overly generic biological terms such as lesion, alongside non-discriminative nouns like role in the context of neutrophils and fibrosis, diluted decision boundaries.
Despite its demonstrated efficacy in refining biomedical text classification, the authors acknowledged notable computational and methodological constraints that govern deployment of the dual-perspective framework. A formal complexity analysis revealed that the integrated pipeline scales according to an asymptotic bound of big O of n squared plus n times d squared plus the cardinality of vertices plus the cardinality of edges plus n cubed, where n represents data points and d represents feature dimensionality. The quadratic demands of HDBSCAN clustering and the cubic upper bound of Support Vector Machine training impose significant computational overhead during initial preprocessing, creating scalability hurdles for massive datasets. Additionally, while Node2Vec effectively captures relational data, it also increases the dataset size and computational burden, potentially limiting scalability in resource-constrained settings. The investigators noted that dataset biases and classifier sensitivity to complex topological manifolds remain inherent limitations. Future research directions outlined by the team include integrating alternative classifiers such as Random Forests and Graph Neural Networks, deploying modern large language models, and extending the methodology to multi-omics workflows.
Subject of Research: Explainable artificial intelligence for inconsistency detection in biomedical text classification
Article Title: Inconsistency detection in cancer data classification using explainable-AI
Article References: Mortezaagha, P., Makarand Joshi, A., & Rahgozar, A. (2025). Inconsistency detection in cancer data classification using explainable-AI. BMC Artificial Intelligence, 1(1), Article 5. https://doi.org/10.1186/s44398-025-00005-6
Image Credits: AI Generated
DOI: 10.1186/s44398-025-00005-6
Keywords: Explainable AI, Cancer Classification, Biomedical NLP, BERTopic, Node2Vec, Knowledge Graphs, Support Vector Machines, Inconsistency Detection, Inconsistency, detection, cancer, data
News Source: Nathaniel Bowman. (November 9, 2026). Explainable AI Framework Cleans Inconsistent Cancer Data to Boost Classification Accuracy. Scienmag.
