Quantum Machine Learning for Medical Prognosis

Introduction

Prognostics and Health Management (PHM) is a crucial aspect of predicting the behavior of machines to support maintenance decisions through diagnosis and prognosis of potential failure modes. In the context of rotating ma chines, diagnosis involves identifying the state of the machine using several traditional Machine Learning (ML) and Deep Learning (DL) techniques.

However, there’s a new and exciting player in town, and its name is Quantum Computing (QC). QC is a rapidly growing field that has contributed to various areas such as optimization, artificial intelligence, simulation, cyberse curity, pharmaceutics, and the energy sector. Although still limited in terms of hardware, QC has been studied as a potential solution for improving models’ computational speed and efficiency.

Enter Quantum Machine Learning (QML), a revolutionary approach proposed in a recent study to diagnose rolling bearings, which are critical components of rotating machinery, using vibration signals. The study used hybrid models involving the encoding and construction of parameterized quantum circuits (PQC) connected to a classical neural network, the Multi-Layer Perceptron (MLP).

Methodology

The step-by-step framework application of QML models is presented in Figure 4. This is a framework based on dif ferent studies that use the PQC logic in QML models. However, in this case, several modifications were implemented regarding the neural network used, PQC settings, and health status diagnosis. Each stage of FIGURE 4 is described below:

• Prepare Quantum Dataset: consists of pre processing the classical data. For example, normalization, dimen sionality reductions, and feature extraction can be performed.

• Evaluate quantum model: after encoding the data, PQC is created. PQC consists of one or several logic gates where the parameters of the gates are free parameters to be adjusted/optimized depending on the error propagated from outside to inside the circuit.

• Sample or Average: measurements are performed, returning the processed quantum data to classical data.

• Evaluate classical model: the classical model is a neural network. Features extracted from the database and encoded are inserted into the neural network. In this process, backpropagation is flowed through the weights of the neural network and into the PQC.

Results

Table IV shows the accuracy, precision, recall, and F1- score results obtained when running the classic ML and QML models. Initially, the lowest accuracy presented is for MLP, resulting in 95.40% and 91.95% for five and eight features, respectively. Among the QML models with five features, the best accuracy (98.08%) is presented in three scenarios: only rotation gates with one layer; and VQE with the CZ gate having one and five circuit layers. However, observing the other metrics, the CZ with five layers had a better-weighted precision than the others.

Considering the first type of QML model (Ry, Rx, Rz), the behavior of the four metrics has better performances on circuits with only one layer and worse ones with ten layers. Thus, showing a decreasing pattern as the number of layers increases. The PQC with CNOT has the lowest accuracy of the QML models when applied to ten circuit layers (95.79%). Increasing this number to 1 peaks the performance to 96.55%. As observed, this result does not improve when increasing the number of layers. The PQC with CZ as the two-qubit gate has its worst result with ten layers and the best with 5 when considering mainly the precision. Finally, the iSWAP has similar behavior to CZ, i.e., the worst result for ten layers. However, best with one and intermediate with ten.

Conclusion

The main objective of this work was to develop a QML- based methodology for the PHM with applications to equip ment through vibration signals. We bring results regarding bearing data available in the literature. The framework of QML models was based on an existing structure, but they have different combinations in the quantum part itself. That is, the proposed PQCs have other configurations in addition to rotation’s gates. These are: the increment of the VQE algorithm combined with different two-qubit gates. Moreover, additional layers of these circuits were used to identify if the increment of quantum operations, such as entanglement, would bring improvements in the results.

Also, it was observed that when compared to the MLP model that has the same neural network configuration used for the quantum models, the QML results were better overall.

Further Reading

This article is based on:

• Prognostics and Health Management of rotating machinery via Quantum Machine Learning - C. Maior et al.

• Quantum Machine Learning: A Review and Current Status - N. Mishra et al.

• Recent advances in quantum machine learning - Y. Zhang et al.

• Quantum machine learning (QML) poised to make a leap in 2023 - Sri Krishna

• 2021 Year in Review - Google Quantum AI