Knowledge And Data Co-driving For Disease Subtype Classification

Non-communicable diseases have rapidly ascended to the forefront as the principal adversaries of human health,becoming the leading causes of mortality.From cancer to cardiovascular diseases,these diseases share a common trait: if identified early with swift and accurate detection,their mortality rates can be significantly reduced.However,the crux of the issue lies in the numerous subtypes of many noncommunicable diseases,which often appear deceptively similar on the surface yet require distinct treatment strategies.This complexity has meant that traditional diagnostic methods heavily rely on the expertise of specialists,leading to inefficiencies.Today,artificial intelligence offers a novel solution to this dilemma.Advanced intelligent models can now diagnose specific subtypes of diseases more accurately,formulating more personalized and effective treatment plans.It is a significant leap in medical technology and brings new hope to countless patients.Artificial intelligence methods to analyze disease subtypes have become a current research focus,transforming from first-generation knowledge-driven models to secondgeneration data-driven models.Initially,model operation relied on expert knowledge and reasoning logic,and despite breakthroughs in diagnosis,it was limited by dependence on expert knowledge.With the advent of big data and cloud computing,second-generation methods utilizing technologies like deep learning have achieved more accurate subtype diagnosis.However,their "black box" nature resulted in diagnostic results lacking logical reasoning and interpretability.Thus,exploring thirdgeneration artificial intelligence methods,which combine knowledge and data to enhance model accuracy and interpretability,is essential.It provides new perspectives and challenges for the field of disease subtype diagnosis:(1)How do we integrate the data-driven model with existing diagnostic knowledge so that experts can empower the model for diagnosis?(2)How can the data-driven model discover unknown diagnostic knowledge and enable the model to assist experts in diagnosis?(3)How can the data-driven model adaptively control diagnostic knowledge so that the model and experts can make interactive diagnoses?Based on these questions,this paper proposes a knowledge and data co-driving approach to design knowledge fusion,discovery,and interaction models for data-driven models,thereby achieving high-precision and interpretable methods for disease subtype classification.The main work is introduced as follows:(1)A data-driven knowledge fusion model(Data & Knowledge,D&K)is proposed to diagnose disease subtypes accurately by injecting currently known and effective diagnostic knowledge into the model.Specifically,it includes an explicit feature extraction module,utilizing multi-instance and ensemble learning to extract and integrate pathological features from different bags.Additionally,based on Gestalt theory in psychology,an expert knowledge fusion module is designed.It maps and measures these features within a three-dimensional knowledge space,then diagnoses them based on their Euclidean distances.Experiments on public and in-house datasets verify D&K’s accuracy and reliable reasoning capabilities in disease subtype diagnosis.(2)A data-driven knowledge discovery model(Multi-Modality Multi-Scale,M3S)is proposed,which enables an accurate and interpretable subtype diagnosis process by autonomously mining potential correlations between data and unknown diagnostic knowledge through the model.Specifically,in the multi-scale feature extraction module,the paper uses Gramian angular field technology to transform spectral data into images of different resolutions,amplifying detail differences and capturing high-dimensional diagnostic features using a dual-branch structure.Additionally,in the multi-modal knowledge discovery module,the model combines preliminary spectral predictions with medical history information,representing the influence of different factors on subtype diagnosis through probability and weight matrices.Experimental tests on in-house datasets show that M3 S outperforms existing advanced models on various evaluation metrics.The learned weight matrices are a basis for subtype diagnosis,forming diagnostic knowledge for medical reference.(3)A data-driven knowledge interaction model(Prompt-Aware KnowledgeTuning,PAKT)is proposed,which can adaptively generate the probability of pathological features with a small number of expert prompts while representing diagnostic knowledge quantitatively as the degree of association between pathological lesions and subtypes,thereby constructing a human-computer knowledge interaction model for subtype diagnosis.Specifically,the prompt-aware module under expert prompts uses a self-supervised pre-trained feature extractor to predict multi-scale pathological feature probabilities adaptively without extensive manual annotation.Meanwhile,the knowledge fine-tuning module quantitatively represents diagnostic knowledge through trainable weight matrices,uncovering the impact and correlations of different pathological probabilities on subtype diagnosis,thus complementing expert knowledge.Extensive experiments on public and in-house datasets show that PAKT maintains high precision in diagnosing disease subtypes while enhancing its work efficiency and interpretability.