💻 computer science

FQXAI: A Semantic and Circuit-Level Explainability Framework for Hybrid Quantum-Classical Medical Learning

This paper introduces FQXAI, a unified framework that enhances the interpretability and robustness of hybrid quantum-classical medical models by combining an exhaustive semantic explanation method (FQ-SHAP) with a noise-aware circuit-level analysis tool (SVQX) to generate human-readable rationales and optimize quantum gate configurations across diverse biomedical datasets.

Original authors: Ankit RAJ, Pramod Parashar, Subramanian Ramalingam

Published 2026-11-10

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ankit RAJ, Pramod Parashar, Subramanian Ramalingam

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the high-stakes world of medical diagnosis, trust is as vital as accuracy. When a computer program suggests a patient has a heart condition or a specific type of cancer, doctors and patients need to know why. For years, powerful artificial intelligence systems have acted like black boxes, delivering correct answers without revealing the logic behind them. This opacity is a major hurdle in healthcare, where understanding the reasoning is often required to ensure safety and align with human biology. Recently, scientists have begun exploring a new frontier: quantum machine learning. These systems use the strange, counterintuitive rules of quantum physics to process information, offering the potential to solve complex medical problems faster than traditional computers. However, these quantum models are even harder to understand than their classical counterparts, often leaving their decision-making processes completely hidden behind layers of abstract math and physical noise.

A team of researchers at Pondicherry University has developed a new framework called FQXAI to solve this problem. Their goal was to create a system that not only predicts medical outcomes using quantum technology but also explains exactly how it reached those conclusions. They built a bridge between the raw, messy data of medical tests and a clear, human-readable logic. Instead of letting the computer work in a hidden space, they forced it to translate complex signals into eight simple, understandable concepts, such as the height of a heartbeat wave or the density of a lung shadow. By doing this, they could generate clear "if-then" statements that describe the model's reasoning in plain language, much like a doctor explaining a diagnosis.

The researchers tested their system on three very different types of medical data: heart rhythm recordings, chest X-rays, and microscopic images of tissue. In each case, the system successfully converted the raw data into these eight core concepts. For example, when looking at a heart rhythm, the model didn't just see a squiggly line; it identified specific features like the "QRS amplitude," which relates to the strength of the heart's electrical signal, and the "R-Peak upstroke," which describes how quickly the signal rises. The system then used fuzzy logic, a method that allows for shades of gray rather than strict yes-or-no answers, to decide if these features were low, normal, or high. This allowed the model to produce a rule like: "If the R-Peak upstroke is high and the QRS amplitude is normal, then the patient likely has a specific type of irregular heartbeat."

To ensure these explanations were not just guesses, the team used a rigorous mathematical method to measure exactly how much each of the eight concepts contributed to the final decision. They calculated the value of every possible combination of these concepts, a task that would be impossible with a larger number of variables but was manageable with their eight-concept limit. This process confirmed that the model was indeed relying on the specific medical features it claimed to use. Furthermore, they traced these insights back to the original images and signals, showing that the model was focusing on the correct parts of a chest X-ray or the right moments in a heart rhythm, just as a human specialist would.

The study also tackled a major physical challenge of quantum computing: noise. Real quantum computers are currently imperfect, and tiny errors in their operations can ruin calculations. The researchers simulated this noise to see how it affected their model. They discovered that some of the quantum operations, which are like the gears in a machine, were actually hurting the model's performance by introducing errors. By identifying and removing these specific, unhelpful operations, they were able to simplify the quantum circuit. This pruning process didn't just make the system smaller; it made it more accurate. In their tests, removing these noisy parts improved the model's accuracy by up to 2.2 percentage points on heart rhythm data, proving that a simpler, cleaner quantum circuit could outperform a more complex, error-prone one.

The results show that it is possible to build hybrid systems that combine classical and quantum computing to diagnose medical conditions while remaining transparent. The framework achieved high accuracy rates, correctly identifying heart issues in about 89 percent of cases and pneumonia in chest X-rays in roughly 77 percent of cases. More importantly, it provided a clear, auditable trail of how these decisions were made. The researchers found that their method was stable even when the simulated environment was noisy, meaning the explanations remained reliable. This work suggests that the future of medical artificial intelligence does not have to be a mystery. By designing systems that translate complex data into simple, logical concepts, scientists can create tools that are not only powerful but also trustworthy enough to be used in real hospitals.

Technical Summary: FQXAI – A Semantic and Circuit-Level Explainability Framework for Hybrid Quantum-Classical Medical Learning

Problem Statement
While hybrid quantum-classical learning models have demonstrated competitive predictive performance in biomedical tasks, their decision-making processes remain opaque. The combination of learned representations, parameterized quantum circuits, entangling operations, and hardware noise creates significant barriers to obtaining human-interpretable explanations. Furthermore, existing Explainable AI (XAI) frameworks often fail to address the specific challenges of Noisy Intermediate-Scale Quantum (NISQ) devices, such as gate infidelities and decoherence, or fail to connect abstract quantum circuit behaviors with clinically relevant semantic concepts. There is a critical gap in frameworks that simultaneously provide semantic rationales (IF–THEN rules) and circuit-level analysis (gate contributions) while accounting for noise.

Methodology
The authors propose FQXAI, a unified framework that bridges high-dimensional biomedical data, intermediate semantic representations, and quantum processing. The framework operates through the transformation $F: X \to \tilde{Z} \to S \to H \to Y$ :

Proxy-Supervised Semantic Interface:
- High-dimensional inputs (ECG, radiology, histopathology) are compressed by a classical feature extractor into an 8-dimensional bottleneck ( $\tilde{Z}$ ).
- This bottleneck is weakly supervised using programmatic proxy targets (e.g., QRS amplitude, alveolar opacity) derived from signal processing and computer vision pipelines, rather than manual annotation.
- A Differentiable Gaussian Fuzziﬁer maps these continuous coordinates into linguistic states ( $K = \{Low, Normal, High\}$ ) using learned membership functions. This enables the generation of model-level IF–THEN rationales.
Quantum Encoding and Processing:
- The semantic states are encoded into an 8-qubit Hilbert space using Angle Encoding ( $R_y$ rotations).
- A variational quantum circuit (ansatz) with $L=2$ entangling layers processes the data. Each layer consists of parameterized single-qubit rotations followed by a chain of Controlled-NOT (CNOT) gates.
- The simulation incorporates a Controlled Depolarizing Noise Model (5% probability) acting on qubits after each CNOT operation to simulate NISQ constraints.
Semantic Attribution (Exact FQ-SHAP):
- Because the semantic interface is constrained to 8 dimensions, the authors compute Exact Shapley Values over all $2^8 = 256$ possible coalitions. This eliminates the need for Monte Carlo sampling approximations.
- These semantic attributions are projected back to the original input space (e.g., ECG time points or image pixels) using a Jacobian-based sensitivity projection, allowing for visual heatmaps of model attention.
Circuit-Level Analysis and Pruning (SVQX):
- The framework adapts the Shapley Values for Quantum Circuit Explanations (SVQX) method. It treats the 14 physical CNOT gates in the baseline circuit as cooperative-game players.
- A class-balanced, target-aligned coalition value function is defined to evaluate the marginal contribution of each gate under the noise model.
- Gates with negative Shapley contributions ( $\Phi_j < 0$ ) are identified as noise-dominant and are pruned. The pruned circuit is retrained under the same noise conditions.

Key Contributions

Proxy-Supervised Semantic Interface: The definition of an 8-dimensional bottleneck with programmatic supervision and differentiable fuzzy membership states, enabling the extraction of human-readable IF–THEN rules that bridge high-dimensional data and logic.
Exact Semantic-Space Attribution: The operationalization of exact Shapley values on the 8-dimensional semantic game, allowing for exhaustive evaluation of all 256 coalitions without sampling variance.
Coupled Representation- and Circuit-Level Analysis: The integration of semantic attribution with SVQX, utilizing a noise-conditioned objective to identify and prune negatively contributing CNOT gates, thereby linking semantic interpretability with noise-aware circuit optimization.

Experimental Results
The framework was evaluated on three biomedical modalities: MIT-BIH Arrhythmia (1D-ECG), PneumoniaMNIST (2D X-Ray), and PathMNIST (2D Histopathology).

Predictive Performance: The hybrid FQXAI models achieved accuracies of 0.890 (ECG), 0.768 (X-Ray), and 0.726 (Pathology), performing competitively against matched classical baselines.
Attribution Fidelity: The framework demonstrated high fidelity in identifying critical features. Exact FQ-SHAP achieved Area Under the Deletion Curve (AUDC) values of 0.780, 0.786, and 0.423 for the respective domains, indicating that ablating top-ranked features rapidly degraded model confidence.
Stability: The semantic rankings showed high robustness to noise, with Pearson correlation coefficients between noiseless and noisy attributions exceeding 0.93 across all domains.
Circuit Pruning: Under 5% simulated depolarizing noise, SVQX-guided pruning reduced CNOT counts from 14 to 11, 12, and 10 for the three domains, respectively. This reduction correlated with accuracy increases of 2.20%, 1.74%, and 1.51%, outperforming matched random-pruning controls.

Significance and Claims
The paper claims that FQXAI demonstrates the feasibility of coupling semantic explanations with noise-aware circuit analysis for auditable hybrid quantum medical learning. By transforming heterogeneous biomedical inputs into a proxy-supervised semantic concept space, the framework provides transparent, model-level rationales (e.g., "IF R-Peak Upstroke is HIGH... THEN Premature Ventricular Contraction"). Furthermore, by treating quantum gates as cooperative players in a noise-conditioned game, the framework enables the identification and removal of detrimental operations, resulting in shallower, more accurate circuits. The authors assert that this approach offers a unified pipeline for examining how learned semantic representations and quantum operations contribute to medical predictions, addressing the "black box" nature of both deep learning and quantum machine learning in high-stakes clinical settings.

Technical Summary: FQXAI – A Semantic and Circuit-Level Explainability Framework for Hybrid Quantum-Classical Medical Learning

More like this