The work addresses a central problem in comparative healthcare: patients treated by different providers may look similar in terms of diagnosis while undergoing substantially different combinations of surgery, axillary management, chemotherapy, radiotherapy and hormonal therapy. Those pathways can influence functional, symptomatic and patient-reported outcomes, so a provider comparison that ignores them may mistake differences in case mix for differences in quality. Value-Based Healthcare seeks to compare patient-centred outcomes in relation to the resources used to achieve them, making comparability essential. Conventional breast cancer classifications often rely on biological or molecular tumor characteristics, whereas this study deliberately based its archetypes on characteristics likely to shape each patient’s clinical trajectory.

The derivation cohort came from the Integrated Health Organization of Ezkerraldea-Enkarterri-Cruces in the Basque Health System, Osakidetza. Eligible patients were adults with newly diagnosed stage I–III invasive breast cancer or ductal carcinoma in situ who received at least one treatment between 2018 and 2020. Rare tumors, lobular carcinoma in situ and metastatic disease were excluded. Baseline and six-month follow-up data included sociodemographic characteristics, clinical outcomes and care-process outcomes aligned with an International Consortium for Health Outcomes Measurement variable set, but only demographic and treatment variables entered the classification models. This design was intentional: patient-centred outcomes were reserved as later benchmarking outcomes, avoiding the circularity that could arise if the same measures used to define the classes also determined their evaluation.

Before the latent models were fitted, clinicians designated patients with ductal carcinoma in situ as a reference class, Arch-0, because of its distinct, generally less burdensome treatment pathway. The remaining complete cases were randomly divided into a training set of 1,256 patients, used to derive the latent structure, and a test set of 314 patients, used to evaluate assignment of new patients. The researchers fitted models containing one through eight classes and compared them with the Akaike information criterion, Bayesian information criterion, sample-size-adjusted BIC, likelihood-ratio and Pearson chi-square statistics, entropy and average posterior membership probabilities. A six-class solution was selected because it combined the lowest adjusted BIC, 9,735.97, with strong separation and clinically recognizable pathways. Its entropy was 0.932, and the average posterior probability across the six classes ranged from 0.92 to 0.98. Examination of bivariate residuals found a maximum value of 1.97 for the association between age and axillary surgery, with no substantial violation of the model’s local-independence assumption identified.

The resulting classes captured clinically distinct combinations of age and treatment. Arch-3, the largest class at 421 patients, and Arch-6, with 162, both featured breast-conserving surgery, sentinel lymph-node biopsy, radiotherapy and hormonal therapy, but differed sharply in chemotherapy use: virtually no Arch-3 patients received chemotherapy, whereas all Arch-6 patients did. Arch-5, containing 368 patients, also largely underwent breast-conserving surgery but had less radiotherapy and no chemotherapy or hormonal therapy in its representative profile. Arch-1 comprised 144 predominantly younger patients who underwent mastectomy with immediate reconstruction, while Arch-2 included 121 mostly older patients treated with mastectomy without immediate reconstruction and more extensive axillary surgery. The smallest class, Arch-4, had 40 patients, 80% of them younger than 50, and received the most intensive combination: mastectomy, axillary clearance, chemotherapy, radiotherapy and hormonal therapy. These are statistical groupings rather than deterministic clinical categories; individual patients within a class do not necessarily share every characteristic of its representative profile.

To test whether the latent model could assign previously unseen patients reproducibly, the investigators used the estimated class probabilities to classify the 314-patient test set. They then trained a Random Forest model on the LCA-assigned archetypes in the training set, optimized it through fivefold cross-validation and applied it to the same test patients. Agreement between the two approaches was 99.68% overall. Sensitivity and specificity were both 100% for Arch-2, Arch-3, Arch-5 and Arch-6, with lower values for Arch-1 and Arch-4. This comparison supports consistency between the LCA-derived labels and the independent machine-learning assignments, but it is not an independent validation of the latent structure itself: the Random Forest was trained to reproduce classes that LCA had already defined. In other words, the exercise evaluated the stability of a classification tool against the pre-established latent classes rather than discovering or confirming those classes without assistance.

Transferability was examined in a separate multicenter dataset from the VOICE community, comprising 352 patients whose classification variables could be entered into the derivation model. The average posterior probability of membership in the most probable class was 0.965, with class-specific averages ranging from 0.913 to 0.997. The overall distributions of age, surgery and treatment variables within the archetypes were broadly similar across cohorts, although several proportions differed. The authors interpreted these differences as potentially reflecting variation in case mix and clinical practice among centers. Importantly, they described the agreement as descriptive and preliminary evidence rather than a formal statistical confirmation of external validity. The external centers also had organizational and population characteristics similar to those of the derivation cohort, so the findings do not establish performance in substantially different health systems or patient populations.

The framework offers a methodological advance over the hierarchical clustering on principal components approach that previously informed the consortium’s work. Clustering requires a subjective choice about where to cut a dendrogram and does not directly provide a probability-based rule for assigning a new patient. LCA instead tests competing class structures against statistical fit criteria and generates posterior membership probabilities, making uncertainty visible rather than hiding it behind a hard group label. That feature could be useful in hospital benchmarking, where an automated classification system must be applied consistently across sites. The approach may also support pathway design and help distinguish groups with different treatment burdens, although the present study did not test whether clinicians can use the archetypes to make better decisions or whether patients experience better outcomes after such use. It also did not directly measure whether provider comparisons based on the classes yield more accurate or actionable rankings.

Several limitations constrain the conclusions. Complete-case analysis excluded 684 patients, or 27.5% of the initial cohort, because age or surgery data were missing. Those excluded patients were older and had lower recorded treatment use, so the missingness cannot be assumed to be random and the resulting classes may underrepresent older, less intensively treated patients. Six months of follow-up may also fail to capture the full trajectory of care for some patients with breast cancer. Class sizes were imbalanced, particularly the 40-patient Arch-4, and individual variation within classes limits the precision of class-level conclusions. The model was derived in one disease, one health system and a relatively short observation period; prospective, temporal and broader external validation, bootstrap class-stability analyses, sensitivity analyses using alternative variable sets, and periodic updating as treatment practices change remain necessary.

Overall, the study presents latent class analysis as a transparent and potentially scalable way to construct therapeutic-pathway archetypes for Value-Based Healthcare. Its strongest demonstrated result is not that a particular hospital performs better, but that a probabilistic six-class framework can reproduce recognizable treatment profiles, assign held-out patients with high agreement against an alternative classifier and produce confident assignments in an independent dataset. The work therefore offers a practical measurement tool rather than evidence of improved clinical decisions or patient benefit. Broader studies will be needed to assess the external validity and stability of the archetypes, determine whether they can support more comparable provider assessments, and evaluate whether the same approach can be transferred to other medical conditions and organizational settings.

Subject of Research: Latent class analysis of breast cancer treatment pathways for healthcare benchmarking

Article Title: From therapeutic pathways to patient archetypes: a latent class analysis framework for Value-Based Healthcare benchmarking

Article References: Otamendi, M., Alayo, I., Gorostiza, A., Cacicedo, J., Fullaondo, A., & García-Lorenzo, B. (2026). From therapeutic pathways to patient archetypes: a latent class analysis framework for Value-Based Healthcare benchmarking. Frontiers in Health Services, 6, Article 1789729. https://doi.org/10.3389/frhs.2026.1789729

Image Credits: AI Generated

DOI: 10.3389/frhs.2026.1789729

Keywords: breast cancer, latent class analysis, patient archetypes, healthcare benchmarking, value-based healthcare, treatment pathways, case mix, patient-centred outcomes, random forest, external validation, therapeutic, pathways