Exploring Quantum kernel methods for breast cancer subtyping: a real-world experiment

Quantum Machine Learning (QML) has been proposed as one of the most promising applications of quantum computing for currently accessible Noisy Intermediate Scale Quantum (NISQ) devices. Quantum kernel estimators, specifically, hold great versatility and can be applied to any classical ML model, both unsupervised and supervised [1]. However, whether these quantum models can outperform classical ML in solving real-world problems still needs to be explored and demostrated [2] as they have so far only been tested on ad hoc simulated data [3] or simplified artificial datasets. We aimed to evaluate the performance and usability of Quantum Support Vector Machine (QSVM) for the classification task of a real complex tumor dataset. The dataset encompasses multi-omics breast cancer sample data and the subtypes are not separable by classical SVM. We therefore tested a QSVM based on ZZ-feature map [3] and compared it to a classical SVM with the same task and optimization procedure. Our results showed that complex quantum decision boundaries suffer from generalization capabilities, in contrast with the ideal simulated dataset for which the used encoding map was formulated. To overcome this limitation, we have explored better-suited feature map configurations, tailored to the data. Our work is one of the first attempts to show how QML can be effectively exploited in a realworld clinical-translational context.