acceptodds
Under review as a conference paper at ICLR 2027

HQAI-TB: A Multimodal Hybrid Quantum–Classical Artificial Intelligence Framework for Tuberculosis Diagnosis from Chest Radiographs and Clinical Records

Abstract

Tuberculosis (TB) is one of the most dangerous and lethal infectious external agents. In highly prone countries, the diagnosis of TB is hindered by three interconnected bottlenecks: low-quality X-rays, lack of expert readers, and clinical information locked away in bilingual or multilingual paper records and not in digital versions. There have been many AI solutions to address these bottlenecks and challenges in silos: chest X-ray classifiers that incorporate clinical information from radiology reports as covariates, document digitization pipelines that largely ignore imaging findings, and hybrid quantum-classical related learning systems that have not yet been tested on TB diagnosis with integrated clinical records. In this paper, we propose a multi-modal framework with four clinically deployable modules, QASP-Net, which conditions the images and enhances on the estimated quality descriptors, called HQAI-TB. VQC-ResNet is a combination, and it works. In VQC-ResNet works and combines. In this paper, we propose HQAI-TB, which is a multi-modal framework consisting of four clinically deployable modules, QASP-Net, which conditions the images and improves the estimated quality descriptors. VQC-ResNet: a hybrid which operates on a residual convolutional encoder and a parameterized variational quantum circuit. TBScript-OCR that converts bilingual or multilingual clinical documents into digital data and finally TB-ICD-BERT that extracts entities and makes it easy for ICD diagnostic coding. A diagnostic probability and a structured clinical summary are output by an attention-weighted fusion layer. We describe the formalism of all the components and analyze the pipeline cost. Final clinical validation will determine ultimate utility to patient care, but we describe a pre-specified validation protocol: corpora are fixed (800 public domain frontal views, 138 Montgomery, 662 Shenzhen), reference standards hold, and five ablation configurations are compared with standard statistical tests. We give one initial quantitative result: TBScript-OCR, initialized from an off-the-shelf transformer-based recognizer, and fine-tuned on a 500-form institutional pilot corpus, recognized approximately of characters ( percent error rate) on the printed bilingual set at one site. We report this as a preliminary finding and note that an character error would be unacceptable for downstream diagnostic coding without the ICD ontology filter, we propose here.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
tuberculosis, hybrid quantum-classical learning, variational quantum circuits, multimodal medical AI, chest radiography, optical character recognition, clinical natural language processing, decision support.

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_HcjLOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes HQAI-TB, a multimodal hybrid quantum–classical framework for TB diagnosis that integrates quality-aware chest radiograph preprocessing, a variational quantum circuit prediction head, bilingual clinical document digitization, and ICD coding via a clinical language model, fused through a presence-gated attention mechanism handling missing modalities. The authors explicitly position this as a system specification and pre-specified evaluation protocol: no end-to-end training has been performed, no paired multimodal patient data exist, no quantum hardware was used, and no diagnostic performance results are reported. The sole quantitative measurement is an 82% character-level OCR accuracy on a 500-form pilot corpus.

Critical strengths

The paper exhibits rare intellectual honesty in a hype-prone domain. It claims no quantum advantage or clinical utility.

Critical weaknesses

This paper lacks of experimantal validation, which is far behind the acceptance bar of top-tier conferences like ICLR.

Questions for authors

/

Reviewer_wi1bOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes HQAI-TB, a multimodal TB diagnosis framework with five parts: 1) QASP-Net: a quality-conditioned mixture of enhancement operators, 2) VQC-ResNet: a ResNet encoder feeding an n-qubit variational quantum classifier, 3) TBScript-OCR: a transformer-based OCR for bilingual scanned forms, 4) TB-ICD-BERT: a transformer-based entity recognizer, and 5) Fusion layer: a gated attention-based missing modalities-aware layer.

The paper is explicitly a specification plus a pre-registered evaluation protocol. It reports no diagnostic results and no quantum-advantage claims.

Critical strengths

Strengths: S1- The authors are candid about what has and has not been done, and they avoid overclaiming quantum advantage.

Critical weaknesses

Weaknesses:

W1- Limited Novelty: The architecture mainly combines existing methods, including ResNet, variational quantum circuits, OCR, BERT, and attention-based fusion. The paper does not establish a sufficiently novel methodological contribution. W2- Insufficient Experimental Results: The complete framework has not been trained or evaluated end-to-end. W3- Dataset Limitations: The imaging and clinical document datasets are not paired at the patient level, and the proposed institutional validation datasets have not yet been collected. W4- Unjustified Quantum Component: The paper does not demonstrate that the quantum classifier outperforms a comparable classical model. Experiments use simulation only, without evidence of practical advantage. W5- Poor Writing Quality: The abstract contains duplicated and incoherent sentences. The manuscript also needs clearer explanations of its contributions and more precise technical descriptions.

Questions for authors

Questions:

Q1- What is the specific methodological novelty beyond combining established architectures and techniques? Q2- Can the authors provide end-to-end diagnostic results on paired, patient-level data? Q3- Does the variational quantum circuit outperform a parameter-matched classical classifier? What theoretical or empirical evidence supports its inclusion?

Suggestions:

Please clearly identify the novel contributions and distinguish them from existing multimodal and hybrid quantum-classical approaches. Provide any available experimental evidence, including baseline comparisons, ablation results, and statistical confidence intervals. Report at least one end-to-end or per-component result on real data. Provide a matched classical MLP comparison at fixed features.

Reviewer_GmfMOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

HQAI-TB proposes quality-aware radiograph preprocessing, a variational quantum classifier, bilingual OCR, clinical coding, and multimodal fusion. The paper specifies a validation protocol. Its only quantitative result is approximately 82% OCR character accuracy from a 500-form institutional pilot; the full system has not been evaluated.

Critical strengths

The paper addresses a practical barrier: combining radiographs with information trapped in paper records. Missing-modality handling is explicitly specified (3.5). The authors clearly separate measured results from planned validation and define component-removal criteria (Table 2). This transparency avoids unsupported performance or quantum-advantage claims.

Critical weaknesses

The main contribution remains a proposal. There are no end-to-end results, paired patient-level multimodal data, or executed component ablations (4.4; Table 3). The submission therefore does not establish that the combined design is useful. The sole OCR result lacks a reported held-out split, baseline comparison, and uncertainty estimate. The proposed correction mechanism is also unevaluated (4.4). This measurement cannot establish a reliable document-processing contribution.

Questions for authors

How many held-out forms and characters produced the 82% result, how were patients/templates separated, and how does the unchanged OCR backbone perform? Is any completed paired-data evaluation or matched classical-head comparison available beyond the supplied manuscript? Planned experiments would not resolve the current evidence gap.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.