Hybrid QML Disease Detection Platform

Executive Summary & Technical Analysis

The Hybrid QML Disease Detection Platform is a well-engineered, full-stack clinical screening workbench that bridges Near-Term Intermediate-Scale Quantum (NISQ) computing with practical healthcare triage.

Key Strengths

  • Methodological Rigor: Enforcing a strict "split-before-fit" architecture via an autonomous LangGraph agent eliminates a major pitfall in automated medical ML pipelines (data leakage).
  • Realistic Hardware Modeling: Benchmarking pure statevector evolution (default.qubit) alongside open-system density matrices (default.mixed) with depolarizing and SPAM noise channels grounds the quantum claims in physical NISQ realities.
  • Practical Clinical Framing: Rather than claiming unrealistic diagnostic accuracy on noisy survey data, the platform smartly positions the 6-Qubit VQC as a high-sensitivity, low-cost preliminary triage filter (τ = 0.35 → >85% sensitivity) to refer at-risk patients for secondary imaging.
  • Parameter Economy: Achieving 74.45% positive recall with just 13 trainable parameters demonstrates the expressive efficiency of parameterized quantum circuits on low-dimensional, collinear datasets.

Critical Observations & Areas for Further Investigation

  1. Classical Baseline Tuning: While default Linear/RBF SVMs collapsed at 6 and 8 features due to class imbalance and collinearity, future benchmarks should include class-weighted SVMs (class_weight='balanced'), calibrated Random Forests, and XGBoost to ensure a completely fair comparison against state-of-the-art tabular baselines.
  2. Statistical Ceiling of Survey Data: The ~55% balanced accuracy ceiling across both classical and quantum models highlights the inherent signal-to-noise limits of self-reported symptom surveys. Integrating continuous biomarker data (e.g., blood panels or spirometry) will better test the true boundary-drawing power of the VQC.
  3. Barren Plateau Mitigation: The observed degradation at 8 qubits (48.44% accuracy) aligns with known barren plateau phenomena in global cost functions. Testing local cost functions (measuring single-qubit observables rather than averaging across all n qubits) or layer-wise training strategies could help scale the architecture past 6 qubits.

Inspiration

Early screening for asymptomatic conditions—such as non-small-cell lung carcinoma and cardiovascular disease—is fundamentally limited by the nature of early-stage medical data: subtle symptoms, high collinearity, subjective survey metrics, and sparse clinical cohorts.

When trained on overlapping, noisy medical indicators, classical maximum-margin classifiers (such as linear SVMs) frequently suffer from catastrophic decision margin collapse, defaulting to trivial majority prediction (100% false-alarm or 0% specificity rates).

Core Research Question: Can parameterized quantum Hilbert spaces uncover non-linear correlations that classical models miss?

By mapping tabular biomedical features into n-qubit quantum states |ψ(x)⟩ ∈ ℂ^(2^n), Variational Quantum Classifiers (VQCs) exploit exponential state spaces and quantum entanglement to construct non-linear decision boundaries with extreme parameter economy. To bridge quantum computing with real-world healthcare, we built an end-to-end clinical screening workbench powered by an autonomous AI Preprocessing Agent and PennyLane NISQ circuit simulators.


What It Does

The Hybrid QML Disease Detection Platform is a full-stack, enterprise-grade research and diagnostic workbench comprising four core subsystems:

1. Autonomous AI Preprocessing Agent

  • Inspects tabular clinical uploads (.csv), infers semantic types, and detects data quality anomalies (missing values, outliers, categorical casing, multicollinearity).
  • Generates an adaptive preprocessing plan with a strict mathematical "split-before-fit" leakage guarantee.
  • Normalizes continuous features into [0, 1] to match quantum rotational gate angles.

2. Hybrid Variational Quantum Classifier (VQC)

  • Y-Axis Angle Embedding: Applies Ry(x_i · π) rotations to initialize product states: text |ψ₀(x)⟩ = ⊗_{i=0}^{n-1} Ry(x_i · π) |0⟩
  • Variational Ansatz: Executes a low-depth parameterized variational circuit with linear or ring CNOT nearest-neighbor entanglement.
  • Hybrid Readout: Measures Pauli-Z expectation values ⟨Z_i⟩, pooled with a trainable classical bias b through a sigmoid activation head: text p̂ = σ( (1/n) ∑_{i=0}^{n-1} ⟨Z_i⟩ + b )

3. Dual Quantum Simulation (Noiseless vs. NISQ Noise)

  • Evaluates pure statevector evolution on default.qubit alongside open-system density matrix simulation on default.mixed.
  • Models physical quantum hardware imperfections via depolarizing Kraus channels (E_depol) and State Preparation and Measurement (SPAM) readout bit-flip noise.

4. Clinical Workbench UI

  • Single Page Application built with React 18, Vite, TypeScript, and Tailwind CSS.
  • Features an interactive Quantum Circuit Designer, dynamic risk stratification (LOW, MEDIUM, HIGH), and operational threshold tuning (τ ∈ [0.1, 0.9]) to optimize sensitivity for clinical triage.

How We Built It

System Architecture & Evaluation Flow

                     Tabular Dataset (Survey / Cleveland)
                                       │
                          ┌────────────┴────────────┐
                          ▼                         ▼
              Classical ML Baselines    Hybrid QML Pipeline (PennyLane)
         • Logistic Regression          • Angle Embedding: RY(x_i * π)
         • Linear / RBF SVM             • Low-Depth Ansatz: 2 Layers (RY + Linear CNOT)
         • 5-Fold Stratified CV         • Measurement: ⟨Z_i⟩ Pooling + Trainable Bias
                          │             • Simulators: default.qubit / default.mixed
                          │                         │
                          └────────────┬────────────┘
                                       ▼
                            Master Evaluation Matrix
         • 6-Qubit VQC: Peak Sensitivity (74.45%, 13 params)
         • Classical RBF SVM: Peak Balanced Acc (54.61%)
         • Classical SVM Collapse at 6/8 Feats (100% FN / 0% Spec)
         • 8-Qubit VQC Degradation (Barren / Diffuse Gradient)
         • Noise Robustness: Stable at 1% Gate/CNOT/Readout Noise

Tech Stack

  • Quantum Computing Layer: Designed with PennyLane 0.38+, parameterized QNodes, Adam optimization, and density-matrix noise channels.
  • Machine Learning & Baselines: Built with Scikit-learn, implementing Linear SVM, RBF SVM with Gaussian kernels, and regularized Logistic Regression benchmarked on stratified holdout test sets.
  • Autonomous Agent Layer: Engineered with LangGraph and LangChain schemas, featuring an isolated runtime store to eliminate data leakage.
  • Backend Architecture: Asynchronous FastAPI service running on Python 3.11 with MongoDB persistence via Beanie ODM, Pydantic v2 validation, Argon2id hashing, and JWT bearer authentication.
  • Frontend Workbench: React 18, TypeScript, Vite, Tailwind CSS, and TanStack Query for client caching and live telemetry monitoring.
  • DevOps & Cloud: Docker multi-stage builds, Nginx reverse proxy, and cloud deployment on Render.

Challenges We Faced

  1. Classical Margin Collapse vs. Quantum Expressibility: On the Survey Lung Cancer dataset (2,998 survey records), symptoms like coughing, wheezing, and shortness of breath exhibit high collinearity (r ≈ 0.40). When scaling to 6 and 8 features, classical Linear and RBF SVMs suffered catastrophic margin collapse (Sensitivity = 100%, Specificity = 0%, predicting positive for all 450 test rows). In contrast, the 6-Qubit VQC maintained balanced decision boundaries, achieving 74.45% positive recall with only 13 parameters.

  2. The "8-Qubit Plateau" & Dimensionality Degeneracy: Expanding from 6 to 8 qubits increased the Hilbert space to 2^8 = 256 dimensions, causing test accuracy to degrade to 48.44% and ROC-AUC to 0.4990. With limited clinical training samples, higher-dimensional state spaces induce barren-plateau-like gradient diffusion. In NISQ algorithms, feature selection budgeting is strictly superior to blind qubit scaling.

  3. Simulating Open-System NISQ Decoherence: Simulating density matrices ρ ∈ ℂ^(2^n × 2^n) requires quadratic memory over statevectors (2^(2n)). We optimized circuit depth to L = 2 layers (2 × (n - 1) CNOTs), keeping memory footprints within lightweight server limits.

  4. Preventing Data Leakage in Automated Pipelines: Preventing subtle distribution leakage required architecting a strict "split-before-fit" protocol: every scaler, imputer, encoder, and Sequential Feature Selector (SFS) fits strictly on X_train, followed by a 7-point validation audit.


What We Learned

  • Extreme Parameter Economy: The 6-Qubit VQC achieved the study's highest clinical sensitivity (74.45%) using just 13 trainable parameters (6 × 2 rotation angles + 1 classical bias), demonstrating that quantum Hilbert spaces can construct complex non-linear decision regions with orders of magnitude fewer weights than classical models.
  • Structural Noise Resilience: Evaluating the 4-qubit VQC under 1% depolarizing gate noise, 2% CNOT noise, and 1.5% readout error yielded virtually identical accuracy (52.67%) and ROC-AUC (0.517 vs. 0.515), demonstrating that low-depth variational circuits possess inherent noise tolerance.
  • Clinical Triage Viability: While survey data places a statistical ceiling on overall balanced accuracy (~55%), tuning the operational decision threshold τ down to 0.35 elevates VQC sensitivity to >85%, making it a viable, low-cost preliminary triage filter to refer at-risk patients for secondary CT imaging.

What's Next

  • Real QPU Benchmarking: Deploy circuits from PennyLane simulators to physical superconducting hardware via IBM Quantum and AWS Braket.
  • Multi-Modal Diagnostic Fusion: Integrate 2D/3D Convolutional Neural Networks for chest CT/X-ray scans fused with tabular quantum feature maps.
  • Quantum Explainability (XAI): Replace heuristic attribution with parameter-shift gradient saliency maps directly derived from quantum circuits.

Built With

Share this project:

Updates

Submission history