Hitkang : A Multi-Modal Deep Learning Framework for Polygenic Risk Assessment and Personalised Clinical Guidance Using CNN-ML Fusion and OCR-Based Biomarker Extraction

This paper presented Hitkang, a multi-modal clinical decision-support platform integrating a lightweight CNN for CT kidney image classification with a Random Forest classifier for 43-feature biomarker and SNP-based disease prediction, unified through a validated weighted probability fusion scheme. The system automates biomarker entry via an EasyOCR-based laboratory report parser and delivers predictions through a secure FastAPI REST backend and a twelve-screen Flutter mobile application, supported by a risk-stratified recommendation engine spanning five disease categories and three risk tiers. The CNN component achieves 99.68 % overall test accuracy with only 2.87 million trainable parameters—outperforming or matching five of the six published CT-based benchmarks surveyed while using 25–94 % fewer parameters than comparable architectures. The weighted fusion approach improves five-class disease prediction macro-F1 from 0.918 (ML-only) to 0.934 (fusion), confirming the complementary discriminative value of imaging and biomarker modalities. The complete system has been described in detail to enable replication, and the CNN architecture has been published in peer-reviewed venues (TIJER Vol. 12, Issue 03, 2025; JETIR Vol. 12, Issue 07, 2025) [12]. Planned extensions include: (a) Grad-CAM saliency map overlays on CT images to provide radiologist-interpretable visual explanations for CNN predictions; (b) federated learning across distributed hospital networks to train updated model versions without pooling raw patient data at a central site; (c) three-dimensional volumetric CT analysis using Vision Transformers to capture inter-slice spatial context lost in single-slice 2D classification; (d) DICOM-protocol integration enabling direct image import from hospital radiology information systems and picture archiving and communication systems (PACS); (e) point-of-care edge deployment on low-power ARM-based devices for primary healthcare settings with limited internet connectivity; and (f) prospective clinical validation studies in collaboration with tertiary care institutions to establish real-world performance on patient populations beyond the benchmark dataset.

[1]  World Health Organization, "Noncommunicable diseases," WHO Fact Sheet, September 2023. [Online]. Available: https://www.who.int/news-room/fact-sheets/detail/noncommunicable-diseases [2]  R. Chauhan, S. Bhatia, and R. Kaur, "Attention-based deep neural model for kidney abnormality classification using CT images," in Proc. 14th Int. Conf. Computing, Communication and Networking Technologies (ICCCNT), IIT Delhi, 2024. [3]  M. Nadeem, K. Alzahrani, and A. Ahmad, "Deep learning-based multiclass kidney stone detection from ultrasound imaging," in Proc. Conf. on Computer Science, Electronics and Computer Systems (CSECS), 2023. [4]  K. S. Kumar, P. Reddy, and R. Rao, "DeepKidney: A multi-class convolutional neural network for CT-based kidney lesion classification," in Proc. 2nd Int. Conf. on Computational Intelligence and Informatics (C2I6), 2023. [5]  P. N. Hamsavath, S. S. Rao, and A. Reddy, "Enhanced deep learning framework for automated kidney CT scan classification," in Proc. Int. Conf. on Key Enabling Cyber Technologies for Societal (CKECS), 2024. [6]  M. S. Hossain, T. Ahmed, and F. Rahman, "A customised convolutional neural network for kidney disease detection from CT images," in Proc. Int. Conf. on Artificial Intelligence in Applications and Power Systems (AICAPS), 2023. [7]  K. Sharma, R. Singh, and A. Juneja, "Hybrid ResNet-101 and custom CNN for CT-based kidney disease classification," arXiv preprint arXiv:2503.XXXXX, 2025. [8]  F. Maqsood, N. Javed, and T. Hussain, "SpinalZFNet: A hybrid multi-branch deep learning model for robust CT kidney abnormality classification," Comput. Biol. Med., 2024. [9]  L. Wang, M. Li, and F. Wu, "Multi-modal deep learning combining CT and MRI for kidney lesion characterisation," Med. Phys., vol. 51, no. 4, pp. 2801–2815, 2024. [10] S. N. Almuayqil, H. Alsalamah, and M. Alshammari, "KidneyNet: An end-to-end segmentation and classification pipeline for CT kidney analysis," Electronics, vol. 13, no. 7, 2024.