Explainable Clustering as a Tool for Lung Neuroendocrine Tumors Classification

Abstract

Introduction Grading and prognostic stratification are central tools in clinical practice: they synthesize complex biological and pathological information into categories, which can be used for risk assessment and decision making. In lung neuroendocrine tumors (NETs), the WHO classification traditionally distinguishes typical (TC) and atypical (AC) carcinoids on the basis of mitotic count (MC) and necrosis. This framework is about to evolve towards the integration of Ki-67, a proliferation index with recognized biological and prognostic relevance. However, Ki-67 role has been highly debated in literature, partly due to the lack of universally accepted cut-offs and inter-observer reproducibility issues. When tumor subgroups are hypothesized, clustering represents a useful tool for identifying homogeneous groups of observations. However, translating groups obtained via unsupervised methods into transparent rules that can be applied to the clinical decision-making process is not straightforward.

Objectives The aim was threefold: (i) to explore an alternative data-driven classification of lung NETs, integrating Ki-67 to MC and necrosis, for identifying clinically meaningful tumor subgroups with added value in grading and prognosis determination with respect to the TC/AC dichotomy; (ii) to translate them into a set of interpretable, reproducible, and applicable decision rules; (iii) to quantify uncertainty associated to cut-offs, providing information about variability and support the evaluation of cases that are close to the decision thresholds.

Methods We applied an explainable clustering strategy, transforming the data-driven groups into explicit classification criteria, to a retrospective cohort of 483 lung NETs. Observations were clustered using KAMILA, applied to necrosis, MC, and Ki-67. Partitions from 2 to 7 clusters were tested; the three-cluster solution was selected based on clinical interpretability and internal validation indices, while the two-cluster solution was compared against the TC/AC benchmark. Then, we fitted conditional inference trees to predict cluster labels from clustering variables, deriving a decision tree. Cut-off stability was assessed through repeated cross-validation, estimating an empirical confidence interval (ECI) for each selected threshold. Survival analyses were used to evaluate the prognostic value of the three-cluster solution for overall survival and relapse-free survival; transferability was explored in an external cohort.

Results Clustering identified three groups, named NET G1 to G3, characterized by progressively higher Ki-67, MC, and frequency of necrosis. The groups only partially overlapped with the TC/AC classification, and provided a better stratification of ACs. Ki-67 resulted as the main discriminating variable, splitting at ≤6% (95% ECI: 4-7%) and >16% (95% ECI: 16-18%), for NET G1 and NET G3, respectively. MC contributed to classification in the intermediate group, while necrosis was not selected in the final system. The resulting groups showed prognostic stratification, particularly for relapse, and the derived decision tree showed transferability to the external cohort.

Conclusion We developed an interpretable, practical, three-tier classification system, which redefines the grading of lung NETs. Results supported Ki-67 integration in lung NETs stratification, confirmed MC as a discriminating factor, and suggested that necrosis might carry lower discriminative information. More broadly, this work supports the wider use of explainable clustering in clinically oriented classification studies, especially when data-driven groups need to be translated into rules that can be inspected, communicated, and applied in practice. Incorporating ECI into the inferential decision tree enables quantification of measurement uncertainty, accounting for inter-observer variability.