When one test is not enough: AI that gives doctors more context to decide
Four MUST innovations explore how AI can connect medical images with physiology, patient history, lab signals, and clinical knowledge, aiding doctors in integrating more relevant evidence into decisions, from disease prediction to critical care.
A doctor rarely bases critical decisions on a single test alone. For example, a scan might detect an abnormality but not clarify its significance, while a lab result could indicate an issue but not specify its location. Similarly, a detail in a patient’s medical history might only become relevant when viewed alongside current symptoms. Doctors must piece these clues together, often under tight time constraints. Many medical AI systems face a similar limitation: they can only analyze what they were designed for. An AI trained on images, for instance, might excel at recognizing visual patterns but lack awareness of the patient’s history, physiology, lab results, or whether their condition is evolving.
Four innovations from Macau University of Science and Technology take a different approach. They explore methods to integrate various types of patient data across conditions like coronary artery disease, glaucoma, and infectious lung disease, enabling AI to offer doctors richer context for diagnoses and decision-making [1]–[4].
When appearance does not tell the whole story
One invention applies this concept to coronary artery disease. Imagine a coronary artery as a road that supplies essential traffic to the heart. A CT scan can identify where the road narrows or is hardened by calcium, while the coronary artery calcium score (CACS) quantifies calcium buildup. However, seeing a constricted road doesn’t necessarily reveal how much traffic is slowed.
The second piece of information addresses this. The invention combines CACS with fractional flow reserve data, known as CT-FFR, which indicates how the narrowing impacts blood flow. A machine-learning model then merges these two insights to estimate the severity of coronary stenosis [1].
In essence, the AI evaluates not just the blockage itself but integrates the artery’s appearance with the functional impact on blood flow, similar to assessing a traffic jam by considering both the size of the obstruction and its effect on traffic movement.
Five machine-learning methods were tested with ten-fold cross-validation. The support vector machine achieved the highest accuracy across all three coronary branches: 68.07% for the left anterior descending artery, 82.47% for the left circumflex artery, and 78.73% for the right coronary artery [1].
Putting the patient behind the eye image
A second invention adapts this concept to glaucoma detection [2]. A retinal photograph can be likened to a close-up snapshot of a small part of a much larger story. The AI initially analyzes the photo, focusing on four key structures that offer crucial insights about the eye: the retinal blood vessels, macula, optic disc, and optic cup. However, even a detailed photograph cannot capture all details about the individual behind it, much like trying to understand a case from a single photo while ignoring the rest of the file.
To address this, the invention incorporates information from the patient’s electronic medical record, including medical history, lifestyle, medications, and prior eye exams. This data is organized into a glaucoma knowledge graph, which is considered alongside features identified within the retinal image [2]. Practically, this approach aims to provide the AI with both the image and additional case information. Instead of relying solely on the eye image, the AI can contextualize visual clues within a comprehensive view of the patient’s overall health.
Seeing the patient as a moving picture, not a snapshot
The two inventions related to infectious lung diseases expand on this concept [3], [4]. When a patient is severely ill, a single test is like a snapshot: it shows what was happening at that moment but doesn’t indicate the future course of the illness. The proposed AI system aims to view more of that ‘movie.’ It combines chest images, clinical notes, lab data, other structured information, and measurements tracked over time. Additionally, it uses a large language model to find connections among various clinical data, while a self-attention mechanism helps determine which information is most relevant when considering all sources together [3].
This highlights the importance of timing. For a patient with a severe lung infection, a single high or low reading can be informative, but understanding the trend, whether signs are improving, stable, or worsening, is even more valuable. A related invention employs Adaptive Fourier Decomposition, a mathematical technique, to identify patterns within daily measurements, aiding the system in depicting how the patient’s condition evolves over time [4].
The goal is to transform this comprehensive view into actionable information for doctors during urgent situations. The pulmonary AI is designed to identify patients at high risk of severe pneumonia and predict outcomes such as ICU admission and mechanical ventilation [3]. It can then combine these predictions with insights from a large language model and an expert knowledge graph to guide treatment choices and resource allocation in critical care.
A more advanced AI necessitates its own evaluation process. The related invention concentrates on gauging the model’s accuracy [4], recommending comparisons between AI predictions and actual patient outcomes. Performance is measured using common metrics such as accuracy, precision, recall, and F1 score. Together, these inventions shift the focus from asking, “What does this patient look like now?” to a more valuable question: “How is this patient changing, and what should doctors anticipate next?”
More context, human judgment at the center
The main innovation isn’t a new AI algorithm but a novel way of applying AI: integrating fragmented health data of a patient before key decisions. These advances in heart, eye, and lung care combine visual info with measurable data – linking anatomy and function, images and medical history, current results and trends, and AI forecasts with clinical judgment. This distinction is important as AI’s role in healthcare expands, aiming not to replace doctors but to offer a more complete view during vital moments. The promise of medical AI isn’t to make decisions for us but to help avoid misinterpreting tests, images, or moments as representative of the whole patient.
From research to real-world application
In addition to their clinical potential, these patented methods could facilitate collaborations with healthcare tech firms, medical imaging providers, digital health developers, and clinical partners. Organizations interested in multimodal medical AI, clinical decision support, or further validation and integration of these technologies are encouraged to explore licensing, joint development, and real-world application opportunities with MUST.
References
[1] P. Luo and Y. Zhang, “一种基于机器学习的冠状动脉狭窄程度预测方法” [A machine-learning-based method for predicting the degree of coronary artery stenosis], China Patent Application Publication CN 117292180 A, Dec. 26, 2023.
[2] K. Zhang, Y. Gao, and Z. Zou, “青光眼辅助预测方法及系统” [Glaucoma-assisted prediction method and system], China Patent Application Publication CN 118711790 A, Sep. 27, 2024.
[3] Z. Han, X. Li, K. Cen, and Z. Liu, “基于多模态的医学算法模型构建系统” [Multimodal-based medical algorithm model construction system], China Patent Application Publication CN 120221056 A, Jun. 27, 2025.
[4] Z. Han, X. Li, K. Cen, and Z. Liu, “基于感染性肺部疾病诊断模型的临床性能验证系统” [Clinical performance validation system based on an infectious pulmonary disease diagnostic model], China Patent Application Publication CN 120221055 A, Jun. 27, 2025.
