Quantum Neural Networks Identify Colorectal Leaks with 83.3% Sensitivity

A new method utilising quantum machine learning improves prediction of anastomotic leaks following colorectal surgery. Vojtěch Novák and colleagues atUniversity of Ostrava show that Quantum Neural Networks, employing ZZFeatureMap encodings, achieve sharply improved sensitivity, reaching 83.3%, in identifying this key, low-prevalence (14%), complication compared to classical machine learning models which achieved 66.7%. The findings highlight the ability of quantum feature spaces to better identify minority classes within clinical data, representing a substantial step towards more accurate risk prediction for patients undergoing colorectal procedures.

Enhanced leak prediction via optimised quantum machine learning configurations

Quantum Neural Networks (QNNs) achieved 83.3% sensitivity in predicting anastomotic leaks following colorectal surgery, a substantial increase over the 66.7% attained by classical machine learning models. Previously, reliably detecting these leaks proved elusive due to their low prevalence, affecting approximately 14% of patients. Anastomotic leaks represent a serious post-operative complication following colorectal resection, significantly increasing morbidity, mortality, and healthcare costs. Accurate pre-operative risk stratification is therefore crucial for optimising patient management, including surgical technique selection and post-operative monitoring protocols. The team at the Technical University of Ostrava discovered that optimised quantum configurations, utilising ZZFeatureMap encodings, better prioritise the identification of these minority class events. This encoding method maps classical data into a quantum Hilbert space, creating a feature space where subtle patterns indicative of leak risk can be more readily distinguished. The ZZFeatureMap specifically employs a combination of Pauli-Z rotations, allowing for complex, non-linear relationships within the data to be captured. Further refinement of the QNNs involved testing various optimizers under simulated noisy conditions, revealing that the EfficientSU2 ansatz paired with the BFGS optimizer achieved an Area Under the Curve (AUC) of 0.809; this performance matched that of the highest-performing classical model, Gaussian Naive Bayes. The EfficientSU2 ansatz is a variational quantum circuit designed for efficient parameterisation and optimisation, while the BFGS optimizer is a quasi-Newton method commonly used in classical machine learning for finding the minimum of a function. A low Brier Score of 0.111, alongside a Log Loss of 0.372 and an Efron’s R2 of 0.129, indicated a strong distributional fit for the optimised quantum configuration. The Brier Score measures the accuracy of probabilistic predictions, with lower values indicating better calibration. Log Loss assesses the performance of a classification model where the prediction input is a probability value between 0 and 1. Efron’s R2 provides a measure of explained variance, similar to the conventional R-squared statistic.

Quantum algorithms enhance detection of rare surgical complications

Ever-finer analytical tools are needed to predict surgical complications such as anastomotic leaks, and this research offers a glimpse of how quantum computing might refine risk assessment. Classical machine learning models often struggle with imbalanced datasets, where the event of interest (in this case, anastomotic leak) is significantly less frequent than the negative cases. This can lead to biased predictions and poor sensitivity. Quantum machine learning, by leveraging the principles of superposition and entanglement, offers the potential to create feature spaces that are more sensitive to these rare events. Encouraging results were achieved with simulated quantum environments, but a key question remains unanswered: can these gains be replicated on actual quantum hardware, which is susceptible to unpredictable errors. Current quantum computers are Noisy Intermediate-Scale Quantum (NISQ) devices, meaning they have a limited number of qubits and are prone to errors caused by decoherence and other factors. The study acknowledges the hurdles posed by ‘noise-induced barren plateaus’, where quantum algorithms become unreliable due to errors, but this does not diminish the significance of the work. Barren plateaus occur when the gradients of the quantum circuit vanish exponentially with the number of qubits, making it difficult to train the model effectively.

Future investigation will focus on exploring error correction techniques to improve the quantum models. Techniques such as quantum error correction and noise mitigation strategies are being actively developed to address these challenges. Quantum feature spaces enabled the models to prioritise detection of this rare event, yielding 83.3% sensitivity compared to 66.7% for traditional machine learning approaches. Translating clinical data into a quantum state using qubits, the fundamental units of quantum information, facilitated this improvement. Each clinical variable is encoded as the state of a qubit, allowing the QNN to explore complex relationships between these variables in a high-dimensional quantum space. Improved sensitivity in detecting these leaks, even within simulations, offers a pathway towards earlier intervention and better patient outcomes for the 14% of patients undergoing colorectal surgery who are at risk. Earlier detection could allow for more proactive post-operative monitoring, such as frequent clinical assessments or imaging studies, and potentially enable timely intervention to prevent the leak from worsening. Furthermore, the development of more accurate risk prediction models could inform shared decision-making between surgeons and patients, allowing for a more personalised approach to surgical planning and post-operative care. The potential for integrating these quantum-enhanced models into clinical workflows represents a significant advancement in the field of surgical risk assessment.

The research demonstrated that Quantum Neural Networks achieved 83.3% sensitivity in predicting anastomotic leaks following colorectal surgery, significantly exceeding the 66.7% of classical methods. This improved sensitivity is important because accurately identifying this rare complication, which affects 14% of patients, is crucial for timely intervention and potentially better outcomes. The study utilised ZZFeatureMap encodings and explored optimisation under simulated noise conditions to achieve these results. Authors plan to investigate error correction techniques to further refine these quantum models for future clinical application.