Background: BC phenotype, including HER2-low, is dynamic during the course of disease. The switch from “0” to “low” HER2 status in metastatic BC represents a highly relevant phenomenon in modern BC oncology, because it allows access to anti-HER2 antibody-drug conjugates, as trastuzumab-deruxtecan. In a previous work, a machine learning model with promising accuracy in predicting HER2-low conversion from primary BC to relapse (Bio1) was developed. A metastatic biopsy represents the standard of care in the management of patients with advanced breast cancer, while further samplings are carried out according to various clinical needs. Therefore, in patients with confirmed HER2-0 status on both primary BC and Bio1, a clinic unmet need persists: identifying who may benefit from a further biopsy (Bio2) to detect later HER2-low gain. Aim of the study: the purpose of this study is to develop a ML model able to detect HER2-low gain in patients with HER2-0 status on both primary BC and the first biopsied metastasis (Bio1). Materials and methods: a large multicenter cohort of BC patients undergoing metastatic tissue sampling were included. In order to develop an explainable supervised ML model, the pipeline included: (1) data preprocessing with imputation, feature scaling and one-hot encoding; (2) feature selection via LASSO regression to retain only informative, non-zero-weight variables; (3) model training using XGBoost with penalization to optimize performance for the minority class (HER2-0 primary - HER2-0 Bio1 - HER2-low Bio2). Results: 1408 patients were included in the study. 247 patients presented HER2-0 status on both primary BC and the first biopsied metastasis (Bio1). In this subgroup of patients, 93 Bio2 were available, of whom 78 samples had a matched HER2 phenotype. At Bio2, the probability to find HER2-low was 30% (69% remained HER2-0, 1% was HER2+). The model has 75% accuracy and 80% sensitivity in the prediction of the acquisition of HER2-low status at Bio2 in patients with consistent HER2-0 status on primary BC and Bio1. Variables associated with the highest importance for the ML model are: time from diagnosis to first biopsy > 24 months, exposure to CDK 4/6 inhibitors, non-visceral metastatic disease. Conclusions: the ML model developed in this work is based on easily accessible clinicopathological features and shows high accuracy and sensitivity in predicting the acquisition of HER2-low phenotype in patients with HER2-0 status on both primary BC and the first biopsied metastasis. The introduction of such a model in clinical practice may support a more rational use of resources by identifying patients most likely to benefit from a further biopsy in terms of therapeutic target gain, thereby promoting equitable access to anti-HER2 ADCs.
Background: il fenotipo del carcinoma mammario, incluso lo stato HER2-low, è dinamico e tende a modificarsi nel corso dell’evoluzione della malattia. Il passaggio da uno stato HER2-0 ad uno stato HER2-low nel setting metastatico rappresenta un fenomeno di grande rilevanza nel campo dell’oncologia moderna, poiché consente l’accesso ai coniugati farmaco-anticorpo anti-HER2, come trastuzumab-deruxtecan. In un precedente lavoro, è stato sviluppato un modello di intelligenza artificiale (machine learning, ML) dotato di un’accuratezza promettente nel predire l’acquisizione del fenotipo HER2-low durante l’evoluzione da carcinoma mammario primitivo a recidiva (Bio1). L’esecuzione di una biopsia metastatica, ad oggi, rappresenta una pratica standard nell’ambito della gestione della paziente con carcinoma mammario avanzato, mentre ulteriori campionamenti sono effettuati in base alle esigenze cliniche. Pertanto, nei pazienti che mantengono uno stato HER2-0 sia alla diagnosi che a livello di Bio1 persiste un bisogno clinico insoddisfatto: identificare chi potrebbe beneficiare di un’ulteriore biopsia metastatica (Bio2) per rilevare un successivo guadagno di HER2-low. Scopo dello studio: lo scopo del presente studio è quello di sviluppare un modello ML in grado di predire la comparsa del fenotipo HER2-low nei pazienti con stato HER2-0 sia a livello del carcinoma mammario primitivo che a livello del primo campionamento metastatico (Bio1). Materiali e metodi: è stata inclusa un’ampia coorte multicentrica di pazienti con carcinoma mammario avanzato sottoposti a biopsia nel setting metastatico. Per sviluppare un modello di ML supervisionato spiegabile, è stato necessario ricorrere a: (1) pre-elaborazione dei dati con imputazione, feature scaling e codifica one-hot; (2) selezione delle caratteristiche tramite regressione LASSO per mantenere solo variabili informative, ovvero con peso diverso da zero; (3) addestramento del modello utilizzando XGBoost con penalizzazione per ottimizzare le prestazioni nella classe di minoranza (HER2-0 primario - HER2-0 Bio1 - HER2-low Bio2). Risultati: 1408 pazienti sono stati inclusi nello studio. 247 pazienti hanno presentato fenotipo HER2-0 stabile da tumore primitivo a Bio1. In tale sottogruppo di pazienti, si è disposto di 93 Bio2, di cui 78 con fenotipo HER2 matchato. La probabilità di riscontrare un fenotipo HER2-low nella seconda biopsia metastatica è risultata essere pari al 30% (il 69% è rimasto HER2-0, l’1% è diventato HER2+). Il modello ha mostrato un’accuratezza del 75% e una sensibilità dell’80% nel predire l’acquisizione del fenotipo HER2-low a livello di Bio2 nei pazienti con stato HER2-0 sia alla diagnosi che a livello di Bio1. Le variabili di maggior rilievo per il modello ML sono: tempo trascorso dalla diagnosi alla prima biopsia metastatica > 24 mesi, esposizione agli inibitori di CDK 4/6, malattia metastatica non viscerale. Conclusioni: il modello ML sviluppato nel presente studio è basato su caratteristiche clinico-patologiche facilmente accessibili e presenta un’elevata accuratezza e sensibilità nel predire l’acquisizione del fenotipo HER2-low in pazienti con stato HER2-0 sia a livello del tumore primitivo che a livello della prima biopsia metastatica. L’introduzione di un simile modello nella pratica clinica può supportare un uso più razionale delle risorse, poiché consente di identificare i pazienti che hanno maggiori probabilità di trarre beneficio da un’ulteriore biopsia nel setting metastatico, promuovendo un accesso equo agli ADCs anti-HER2.
Sviluppo di un modello di intelligenza artificiale per la predizione dell'acquisizione del fenotipo HER2-low nelle metastasi da carcinoma mammario
DALLE MULE, ILARIA
2024/2025
Abstract
Background: BC phenotype, including HER2-low, is dynamic during the course of disease. The switch from “0” to “low” HER2 status in metastatic BC represents a highly relevant phenomenon in modern BC oncology, because it allows access to anti-HER2 antibody-drug conjugates, as trastuzumab-deruxtecan. In a previous work, a machine learning model with promising accuracy in predicting HER2-low conversion from primary BC to relapse (Bio1) was developed. A metastatic biopsy represents the standard of care in the management of patients with advanced breast cancer, while further samplings are carried out according to various clinical needs. Therefore, in patients with confirmed HER2-0 status on both primary BC and Bio1, a clinic unmet need persists: identifying who may benefit from a further biopsy (Bio2) to detect later HER2-low gain. Aim of the study: the purpose of this study is to develop a ML model able to detect HER2-low gain in patients with HER2-0 status on both primary BC and the first biopsied metastasis (Bio1). Materials and methods: a large multicenter cohort of BC patients undergoing metastatic tissue sampling were included. In order to develop an explainable supervised ML model, the pipeline included: (1) data preprocessing with imputation, feature scaling and one-hot encoding; (2) feature selection via LASSO regression to retain only informative, non-zero-weight variables; (3) model training using XGBoost with penalization to optimize performance for the minority class (HER2-0 primary - HER2-0 Bio1 - HER2-low Bio2). Results: 1408 patients were included in the study. 247 patients presented HER2-0 status on both primary BC and the first biopsied metastasis (Bio1). In this subgroup of patients, 93 Bio2 were available, of whom 78 samples had a matched HER2 phenotype. At Bio2, the probability to find HER2-low was 30% (69% remained HER2-0, 1% was HER2+). The model has 75% accuracy and 80% sensitivity in the prediction of the acquisition of HER2-low status at Bio2 in patients with consistent HER2-0 status on primary BC and Bio1. Variables associated with the highest importance for the ML model are: time from diagnosis to first biopsy > 24 months, exposure to CDK 4/6 inhibitors, non-visceral metastatic disease. Conclusions: the ML model developed in this work is based on easily accessible clinicopathological features and shows high accuracy and sensitivity in predicting the acquisition of HER2-low phenotype in patients with HER2-0 status on both primary BC and the first biopsied metastasis. The introduction of such a model in clinical practice may support a more rational use of resources by identifying patients most likely to benefit from a further biopsy in terms of therapeutic target gain, thereby promoting equitable access to anti-HER2 ADCs.| File | Dimensione | Formato | |
|---|---|---|---|
|
Tesi di laurea - Ilaria Dalle Mule.pdf
Accesso riservato
Dimensione
1.6 MB
Formato
Adobe PDF
|
1.6 MB | Adobe PDF |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/87288