Artificial intelligence system and software as a medical device for multi-biomarker quantification in clinical diagnosis

Abstract

An AI-powered diagnostic and prognostic system is disclosed, architected as a modular Software as a Medical Device. The system fuses patient-specific data with canonical medical knowledge to power a Recommendation Module that dynamically recommends an optimal set of diagnostic tests. An Interpretation Module then processes these multi-modal inputs, generating a single, quantifiable diagnostic metric and a transparent explanation of the result. A Prognosis Module analyzes the metric over time to forecast disease progression, providing actionable clinical outputs. The entire workflow is underpinned by a robust Compliance Module ensuring safety and regulatory adherence. A continuous Learning Loop feeds clinical outcomes back into the system, iteratively refining its AI models. This creates a holistic, trustworthy, and self-improving platform that transforms raw data into a clear diagnostic and prognostic pathway, enhancing accuracy and patient care.

Description

Artificial Intelligence System and Software As A Medical Device For Multi-Biomarker Quantification In Clinical Diagnosis

[0001] The present invention relates to the field of medical diagnostics. More particularly, it provides an Artificial Intelligence (AI) based system and method, architected as a modular Software as a Medical Device (SaMD), for clinical diagnosis and prognosis. The system dynamically selects and integrates a plurality of biomarkers to generate a quantifiable, explainable, and prognostically useful diagnostic metric, while ensuring rigorous regulatory compliance and patient safety. This invention is especially relevant to point-of-care and resource-limited settings, aligning with global healthcare criteria for accessible diagnostics.

[0002] Medical diagnosis is fundamentally an exercise in synthesizing complex information. Clinicians routinely consider a diverse array of data points – patient symptoms, laboratory results, imaging findings, genetic markers, and other biomarkers – each providing a partial view of a patient's health. No single test or biomarker is typically sufficient for a definitive diagnosis, which is why multiple diagnostic tests are often used in combination. This fragmented approach places a significant cognitive burden on healthcare providers, who must mentally integrate disparate results in a process that is subjective, varies between practitioners, and is difficult to standardize. The challenge is not merely aggregating data, but understanding the often non-linear interactions between different indicators of disease.

[0003] In recent times, the field of medical diagnostics has witnessed significant advancements. Traditionally, medical conditions and diseases are diagnosed using a single diagnostic test. However, no diagnostic test is entirely reliable, and often, a single test is insufficient to determine the presence or absence of a disease. Clinicians, therefore, frequently use multiple diagnostic tests administered either in parallel or in series to enhance diagnostic accuracy. The synthesis of the results of these tests falls upon the clinician, who must interpret the collective outcomes to determine the presence or absence of a disease, as well as the degree of disease progression or remission. This process requires significant clinical judgment and expertise. According to a study published in the Journal of Critical Care Medicine, most diagnostic tests are imperfect, necessitating the use of multiple tests to improve diagnostic certainty, and clinicians are tasked with synthesizing these results to make informed decisions about disease presence and progression (Journal of Critical Care Medicine, 2020, Vol 7(3), pp. 241-248).

[0004] The concept of biomarkers plays a crucial role in the diagnosis and management of diseases. According to the National Institutes of Health, a Biomarker is a defined characteristic that is measured as an indicator of normal biological processes, pathogenic processes, or responses to an exposure or intervention, including therapeutic interventions. Biomarkers can be molecular, histologic, radiographic, or physiologic characteristics. They are not assessments of how a patient feels, functions, or survives but are objectively measured and evaluated as indicators of these processes. The BEST (Biomarkers, Endpoints, and other Tools) Resource, developed by the FDA and NIH, harmonizes terminology and uses of biomarkers and endpoints, facilitating the progression from basic biomedical research to medical product development and clinical care. The use of consistent and mutually understood terminology helps accelerate the development, validation, and qualification of medical product development tools (BEST Resource Fact Sheet, 2024, pp. 2-5; NIH Biomarker - Bookshelf, 2024, pp. 1-3).

[0005] The relatively low cost of diagnostic testing as a fraction of total health care spending belies the crucial role played by testing in determining health outcomes. Cutting-edge diagnostic techniques, such as multivariate index assays based on genomic and proteomic analysis, and new modes of delivery, such as home-based testing kits for sexually transmitted diseases (STDs) and rapid diagnostic tests for other infectious diseases, promise faster, more accurate, and more comprehensive diagnoses of diseases. However, a slow and maladapted regulatory approval process, uncertainties regarding coverage decisions and remuneration, and clinical resistance to new diagnostic methods combine to stifle innovation and slow the dissemination of advanced diagnostic technologies into mainstream clinical use. According to a study in Lab Medicine, diagnostic tests, although they comprise less than 5% of hospital costs, influence as much as 60%-70% of healthcare decision-making (Lab Medicine, 2007, Vol 38(4), pp. 132-136).

[0006] The accuracy of diagnostic tests is not always guaranteed, and this poses a significant challenge. Various biases and sources of variation influence the accuracy of diagnostic tests, including the calculation and interpretation of test characteristics such as sensitivity, specificity, positive predictive value, and negative predictive value. The need for multiple tests and the variability in their results often complicate the diagnostic process. According to a study in the Journal of Critical Care Medicine, the accuracy of a diagnostic test is determined by comparing the test's results against a reference standard, which is considered more accurate. This process requires careful interpretation of numerical test accuracy metrics to account for test errors (Journal of Critical Care Medicine, 2021, Vol 7(3), pp. 241-248).

[0007] The cost-effectiveness of diagnostic tests is another important consideration for their use in healthcare systems. Decision-makers must evaluate the relative costs and outcomes of diagnostic testing to determine when a test is worth its cost. According to the Journal of Clinical Microbiology, molecular diagnostic tests, although more expensive, often lead to significant cost savings by reducing the length of hospital stays and the use of antibiotics, which offsets the additional cost of the test. However, many institutions have a "siloed" approach to budgeting, which leads them to consider only laboratory costs and view novel tests as expensive compared to traditional methods (Journal of Clinical Microbiology, 2010, Vol 48, pp. 3236-3243).

[0008] The diagnostic landscape is further complicated by the need for sophisticated laboratory infrastructure and trained personnel, which are often unavailable in remote or resource-limited settings. According to a discussion published in BMJ Global Health, participants noted that the effective use of diagnostics is hindered by challenges such as cost, staff capacities, irregular supply, transportation delays, inconsistent funding, and poor forecasting and stock management. These operational and health system challenges far outweigh the technical characteristics of the diagnostics themselves (BMJ Global Health, 2016, Vol 1, Article e000132).

[0009] The World Health Organization (WHO) has outlined the ASSURED criteria for diagnostic tests, emphasizing the need for tests to be Affordable, Sensitive, Specific, User-friendly, Rapid and robust, Equipment-free, and Deliverable to end-users. These criteria are crucial for ensuring that diagnostic tests are accessible and effective, particularly in resource-limited settings. According to the WHO report on Mapping the landscape of diagnostics for sexually transmitted infections, the ASSURED criteria are essential for developing and implementing diagnostic tests that can address the unique challenges faced by healthcare providers in diverse settings (World Health Organization, 2004. No. TDR / STI / IDE / 04.1.).

[0010] Artificial Intelligence (AI) has revolutionized the field of medical diagnostics by enhancing the accuracy, speed, and efficiency of the diagnostic process. AI models can analyze vast amounts of data, including medical images, bio-signals, vital signs, demographic information, medical history, and laboratory test results. This comprehensive data analysis supports decision-making and provides accurate prediction results, helping healthcare providers make more informed decisions about patient care. AI integration of multiple data sources offers a more holistic view of a patient’s health, reducing the chance of misdiagnosis and improving diagnostic accuracy. The combination of multimodal data allows for better monitoring of disease progression and management of chronic conditions. AI can also recommend the most appropriate diagnostic tests for individual patients based on their unique clinical profiles, ensuring personalized and precise care. According to a study published in the journal Diagnostics, AI-based diagnostic tools can significantly improve medical diagnostics by analyzing multimodal medical data and providing real-time clinical decision support (Diagnostics, 2023, Vol 13, Article 688).

[0011] Despite its potential, the implementation of AI in medical diagnostics faces several challenges. The integration of multiple modalities into objective criteria and metrics using advanced AI methods requires domain-specific expertise and interdisciplinary skills, which are currently lacking. Additionally, AI systems must adhere to the ASSURED criteria, ensuring that diagnostic devices are Affordable, Sensitive, Specific, User-friendly, Rapid and robust, Equipment-free, and Deliverable to end-users. Meeting these criteria is essential for effective implementation in resource-limited settings. As noted in the United States Government Accountability Office report on Artificial Intelligence in Health Care, while AI has the potential to address many diagnostic challenges, the successful deployment of AI technologies requires rigorous clinical validation and a collaborative approach involving developers, providers, and regulators (GAO, Artificial Intelligence in Health Care. GAO-22-104629: Sep. 2022).

[0012] The integration of artificial intelligence (AI) in medical diagnostics presents significant regulatory challenges, particularly in terms of compliance with medical device regulations. In India, the Medical Device Rules 2017, framed under the Drugs and Cosmetics Act, 1940, outline stringent requirements for the approval and marketing of medical devices. These regulations mandate rigorous clinical validation, quality assurance, and compliance with established standards to ensure the safety and efficacy of medical devices used in patient care. One of the critical challenges with current AI methods in medical diagnostics is their alignment with medical device regulations. AI algorithms must undergo extensive clinical validation to demonstrate their accuracy, reliability, and clinical utility. This involves a series of steps, including preclinical studies, clinical trials, and post-market surveillance, to ensure that the AI system performs consistently and safely in real-world clinical settings. Many AI-based diagnostic tools face hurdles in meeting these regulatory requirements, primarily due to the lack of comprehensive clinical validation and the dynamic nature of AI models that continuously learn and evolve.

[0013] Patent Application WO2024192175A1 discloses a multimodal AI system for diagnostic and prognostic prediction using medical imaging and clinical variables; however, this system does not integrate Large Language Models (LLMs) or federated knowledge graphs and lacks a mechanism for the dynamic, patient-specific biomarker selection and closed-loop learning that are central to the present invention.

[0014] U.S. Patent Application Publication US20220292674A1 discloses what is a sophisticated data fusion technique termed "deep orthogonal fusion." It uses advanced AI (deep learning) to combine highly diverse data types, including radiomics, genomics, and clinical data, for cancer prognosis. It functions by how it mathematically projects these different data modalities into a shared space to identify prognostic patterns. However, a crucial inventive gap exists because this powerful technique operates as a statistical "black box." The method makes no use of canonical clinical knowledge graphs to provide context or clinical plausibility to the patterns it discovers. It lacks dynamic personalization features, as the fusion model is trained once and then applied statically, and it entirely omits the integrated regulatory compliance framework required to ensure the safety, reliability, and auditability of a SaMD.

[0015] U.S. Patent Application Publication US20240386015A1 describes what is fundamentally a safety and reliability framework for artificial intelligence. It correctly identifies the critical risk of AI-generated "hallucinations" (factually incorrect but plausible-sounding outputs) and proposes a system for regulatory oversight in composite AI architectures. It functions by how it polices the output of an AI model, acting as a supervisory layer to detect inconsistencies. However, a significant inventive gap exists because this system is a reactive, safety-centric overlay, not a proactive diagnostic engine itself. It fails to teach the core inventive functionality of the present invention. Specifically, it does not disclose a method for biomarker personalization, wherein diagnostic tests are dynamically selected for an individual patient's unique profile. Its architecture is generic for "composite AI" and is not a modular Software as a Medical Device (SaMD) structure specifically architected for the clinical diagnostic workflow.

[0016] Chinese Patent CN114710707A discloses what is a composite diagnostic score specifically for liver cancer. It embodies the general concept of using AI to combine multiple biomarkers into a single, actionable score. It functions by how it applies an AI algorithm to a predefined set of liver-related biomarkers. The inventive gap is substantial due to the static and inflexible nature of this approach. The system is limited to a fixed, pre-selected panel of markers, making it a "one-size-fits-all" solution for a single disease. It cannot adapt if a new, more effective biomarker becomes available, nor can it deviate from its panel if a specific patient's clinical picture suggests an alternative test would be more informative. This is in direct contrast to the present invention's dynamic, personalized test selection. The Chinese patent also lacks any teaching of adaptive weighting, where a biomarker's contribution to the score can be modified based on the patient's broader clinical context, and it provides no capability for prognostic forecasting, limiting its utility to a single diagnostic snapshot.

[0017] Similarly, Chinese Patent CN111210366A describes what is a composite tumor index system which works by how it aggregates various tumor biomarkers into a single severity index. It demonstrates the appeal of consolidating complex biomarker data into one metric. However, it possesses a critical inventive gap in its lack of contextual intelligence and clinical integration. It provides no mechanism for the integration of canonical clinical knowledge; the system is a "data-in, number-out" calculator with no understanding of established practice guidelines from authoritative bodies or the latest medical research. It cannot, for example, evaluate a biomarker's significance in the context of a patient's comorbidities or treatment history. Furthermore, it lacks an adaptive scoring mechanism that can learn and improve from clinical outcomes and is devoid of any compliance monitoring framework, rendering it unsuitable for deployment as a regulated and trustworthy SaMD.

[0018] Chinese Patent CN109939901A teaches what is a composite index designed for the specific task of predicting Alzheimer’s disease progression. It functions by how it aggregates known biomarkers associated with the disease. Why it is relevant is its application of multi-biomarker aggregation to a prognostic, rather than purely diagnostic, problem. Its inventive gap lies in its lack of architectural sophistication and practical deployability for general diagnostics. It does not teach the architectural modularity of the present invention, where functional components (recommendation, interpretation, prognosis) are designed as independent, validatable units. This monolithic design makes it difficult to update or adapt. Moreover, it provides no framework for ensuring compliance with WHO-ASSURED criteria, limiting its applicability in global health settings, and it lacks the capacity for the real-time, patient-personalized analysis that is a cornerstone of the present invention's dynamic methodology.

[0019] Chinese Patent CN110925640A describes what is a cardiovascular risk score derived from multi-omics data. It integrates complex biological data into a single risk metric. The inventive gap is defined by its static and opaque nature. The system does not support longitudinal or adaptive learning over time; it is a snapshot model that cannot improve its predictive accuracy as it encounters new patient cases and validated outcomes. It also lacks modular explainability, making it difficult for a clinician to understand the basis of its risk assessment, and is missing the integrated compliance monitoring features (e.g., audit trails, risk mitigation) that are core aspects of a safely deployable, regulated SaMD.

[0020] Chinese Patent CN113520787A discloses what is a perioperative monitoring system that functions by how it combines AI with data from wearable devices. It uses AI and biomarker data in a real-time monitoring context. The inventive gap is its profound lack of generalizability. The application is purpose-built for a single, specific clinical scenario (perioperative recovery). It lacks the disease-agnostic platform architecture of the present invention, which is designed to be extensible to a vast range of infectious and non-communicable diseases. Furthermore, it does not teach the dual-layer explainability or the deep, LLM-based knowledge integration that are necessary to tackle complex, general diagnostic challenges beyond simple monitoring.

[0021] Canadian Patent CA2650872C discloses what is a method for selecting relevant biomarkers from a plurality and generating a statistical model that results in a composite score. It touches upon the core ideas of biomarker selection and combination. However, it has a critical inventive gap in its methodology. The method relies on a static, pre-selected set of biomarkers for a given condition. It does not teach or suggest the dynamic, AI-driven selection of the most appropriate biomarkers tailored to an individual patient's unique and evolving clinical profile, which is a core feature of the present invention that moves diagnostic practice from generalized panels to truly personalized medicine.

[0022] U.S. Patent Application Publication US20180068083A1 describes what is a machine learning classifier based on a fixed panel of at least two biomarkers and clinical parameters. Why it is relevant is its use of machine learning for classification based on combined inputs. The inventive gap is twofold: it lacks patient-specific biomarker recommendation, instead relying on a fixed input panel, and it omits the broader integration of medical knowledge taught by the present invention. It functions as an isolated classifier on a predefined set of inputs, in contrast to the present invention which intelligently selects its inputs and then interprets them in the context of a vast, structured medical knowledge base.

[0023] U.S. Patent US20040107124A1 discusses what is a software framework using ontologies to manage regulatory compliance. Why it is relevant is its acknowledgment of the need for structured compliance in software. The inventive gap exists because this provides only a general approach and is not specifically architected for the unique lifecycle and inherent risks of an AI-driven, adaptive SaMD. A generic ontology framework does not address AI-specific failure modes like model drift, data bias, or LLM hallucination. The present invention's compliance module, in contrast, is purpose-built to manage these unique AI risks within a medical device context.

[0024] U.S. Patent US11367184B2 discloses what is a platform for explainable AI (XAI) in the field of digital pathology, which functions by how it allows users to query the AI’s reasoning, addressing the "black box" problem of AI. The inventive gap remains substantial because it is limited to a single layer of explanation: technical attribution (e.g., highlighting the pixels on a slide that were most influential). It does not teach the present invention's dual-layer explainability model, which critically adds a second layer: a plain-language rationale that is validated against a medical knowledge graph. This second layer explains the clinical significance of the technical findings and includes integrated mechanisms for detecting and managing AI-generated errors or hallucinations within that explanation, a safety feature the '184 patent lacks.

[0025] A significant and overarching inventive gap exists in the prior art concerning the creation of truly adaptive, continuously learning diagnostic systems. No prior art teaches a general, closed-loop system that captures real-world clinical outcomes and feeds that information back into an AI diagnostic model to iteratively and safely refine every operational parameter, from biomarker selection logic and weighting schemes to prognostic forecasting models, all within a robust, regulatory-compliant framework.

[0026] The present invention discloses a computer-implemented system and method for AI-assisted clinical diagnosis and prognosis, termed ASSURED-AID (Affordable, Sensitive, Specific, User-friendly, Rapid & robust, Equipment-free, Deliverable - Artificial Intelligence-assisted Diagnostic device). Implemented as a modular Software as a Medical Device (SaMD), the system is architected to holistically address the shortcomings of existing diagnostic paradigms by providing a unified platform that performs end-to-end diagnostic reasoning in a transparent, compliant, and continuously improving manner.

[0027] The system's operation begins with a Recommendation Module that dynamically recommends an optimal, patient-specific set of diagnostic tests. Based on a patient's initial clinical profile, including symptoms and history, this module uses an AI-driven decision engine to select the most efficient diagnostic pathway, moving beyond static testing panels to a personalized approach.

[0028] Following the recommended testing, an Interpretation Module ingests and aggregates the resulting multi-modal biomarker data. This includes, but is not limited to, laboratory results, imaging data, and textual clinical notes. In a key inventive step, the system performs a federated synthesis, fusing this patient-specific data with a vast repository of canonical medical knowledge. This knowledge, extracted from authoritative sources like clinical practice guidelines and medical literature using advanced AI such as Large Language Models (LLMs), provides essential context that allows the system to interpret biomarker values not in isolation, but in light of the patient's complete clinical picture.

[0029] From this enriched and contextualized data set, the Interpretation Module's engine computes a single, quantifiable composite diagnostic metric. This score objectively reflects the likelihood or severity of the patient's condition based on the totality of the evidence. To ensure clinical trust and utility, this metric is presented with a novel dual-layer explanation. This explanation consists of: first, a technical attribution that details the contribution of each biomarker to the final score, and second, a plain-language, fact-checked rationale that explains the clinical reasoning behind the conclusion.

[0030] The utility of the system extends beyond diagnosis through a Prognosis Module. This module analyzes the trajectory of the composite diagnostic metric over time to forecast disease progression, treatment response, or the risk of future adverse events, providing clinicians with a powerful, forward-looking tool.

[0031] The entire system is governed by foundational architectural principles. It is architected to be compliant by design, with an integrated Compliance Module ensuring adherence to medical device standards (e.g., IEC 62304 for software lifecycle, ISO 14971 for risk management) and applicable data privacy regulations. A closed-loop learning mechanism allows the platform to continuously and safely improve its algorithmic performance by incorporating validated clinical outcomes from real-world use. Finally, the system's disease-agnostic architecture and adherence to WHO-ASSURED criteria make it a uniquely extensible and accessible tool, capable of being deployed across diverse clinical specialties and healthcare environments.

[0032] The prior art faces a significant and multifaceted technical problem in creating a truly holistic, adaptive, and trustworthy AI-driven diagnostic system for clinical use. Existing solutions are fragmented and fail to provide a unified platform that addresses the complete diagnostic lifecycle in a regulated environment. A significant gap exists in moving beyond static test panels to a dynamic, personalized approach where biomarker selection is optimized for each patient. Furthermore, prior systems lack a robust mechanism for integrating raw patient data with the vast repository of canonical medical knowledge (e.g., clinical guidelines, research) in a way that is structured, verifiable, and directly influences the diagnostic calculation. A critical technical challenge remains in generating not just a diagnostic score, but a comprehensive output that is both quantifiable and transparent, providing clinicians with a dual-layer explanation that covers both the technical data contributions and the clinical rationale, while actively detecting and mitigating the risk of AI-generated misinformation ("hallucinations"). Finally, no known system provides a general, closed-loop architecture that continuously learns from real-world clinical outcomes to safely refine all operational parameters within an integrated, regulatory-compliant framework designed as a Software as a Medical Device (SaMD)..

[0033] Therefore, a significant gap exists in the art for a AI-driven diagnostic platform that holistically combines: (1) dynamic patient-specific biomarker selection; (2) integration of curated medical knowledge (e.g., guidelines) with patient data via LLMs; (3) generation of a single quantifiable composite diagnostic metric from multiple inputs; (4) AI-driven prognostic forecasting based on the temporal evolution of that metric; (5) a dual-layer explainability framework (technical attribution + plain-language rationale) with safeguards against AI errors; (6) a modular SaMD architecture designed for automated regulatory compliance and risk management; and (7) a closed-loop learning mechanism that continuously refines the system based on clinical outcomes.

[0034] The present invention solves the aforementioned technical problem by providing a comprehensive, computer-implemented system and method for AI-assisted clinical diagnosis and prognosis, termed ASSURED-AID (Affordable, Sensitive, Specific, User-friendly, Rapid & robust, Equipment-free, Deliverable - Artificial Intelligence-assisted Diagnostic device) that has the following specific objectives:

[0035] A primary objective is to move beyond static test panels by enabling dynamic multi-biomarker selection, wherein the system is configured to intelligently select the most relevant biomarkers or diagnostic tests for an individual patient based on their specific clinical profile and context, ensuring personalized, case-specific diagnostic strategies.

[0036] Another objective is to achieve multi-modal data integration, wherein the system is configured to integrate a wide variety of biomarkers (e.g., molecular, physiologic, radiographic, biochemical, digital) and clinical data into a single unified analytical framework, analyzing these heterogeneous inputs together to reflect a holistic view of the patient's condition.

[0037] A further objective is the incorporation of medical knowledge via AI, wherein the system is configured to leverage advanced AI methods, including Large Language Models (LLMs) and federated knowledge graphs, to incorporate canonical medical knowledge (such as clinical guidelines, medical research, and epidemiological data) into the diagnostic process to contextualize raw patient data.

[0038] It is an objective of the invention to generate a quantifiable diagnostic metric, translating the multitude of input data into a single quantifiable diagnostic score or index that objectively represents the likelihood of a disease or health status, calculated through transparent and clinically validatable algorithms.

[0039] It is another objective to provide prognostic forecasting by embedding capabilities to analyze the diagnostic metric over time, forecasting disease progression or remission trajectories and thereby extending the system's utility from diagnosis to prognosis.

[0040] A key objective is to ensure explainability and clinician trust through a dual-layer explainability framework.

[0041] An important objective is to achieve regulatory compliance by design, architecting the system as a SaMD that inherently meets medical device regulatory requirements.

[0042] It is a further objective to implement continuous learning and improvement via a closed-loop feedback mechanism whereby verified clinical outcomes are fed back into the AI models to refine diagnostic algorithms, biomarker weightings, or threshold values, under strict safety and validation checks.

[0043] Yet another objective is a modular and scalable architecture, dividing the system into distinct functional modules with well-defined interfaces to facilitate scalability, maintenance, and disease-specific customization.

[0044] A further objective is seamless hardware integration, ensuring the software platform can be integrated with physical diagnostic devices (e.g., imaging equipment, point-of-care analyzers) and healthcare hardware, with support for standard data exchange protocols.

[0045] It is another objective to ensure ASSURED criteria compliance, meeting the World Health Organization criteria for diagnostics to be Affordable, Sensitive, Specific, User-friendly, Rapid and robust, Equipment-free (or minimal), and Deliverable to end-users, making it practically deployable in resource-constrained settings.

[0046] A final overarching objective is disease-agnostic extensibility, creating a platform whose underlying architecture is flexible enough to be configured or trained for a wide range of medical conditions by updating disease-specific models or parameters without redesigning the entire system.

[0047]

[0048] The present invention provides numerous advantageous effects over the prior art. A primary advantage is Dynamic Multi-Biomarker Selection, moving beyond static test panels to enable intelligent, case-specific selection of the most relevant biomarkers for an individual patient. This ensures a personalized diagnostic strategy that optimizes resource use and diagnostic yield.

[0049] Another advantage is Holistic Multi-Modal Data Integration, wherein the system integrates a wide variety of biomarkers and clinical data into a single unified analytical framework, reflecting a holistic view of the patient's condition.

[0050] A further advantage is the Incorporation of Medical Knowledge via AI, leveraging LLMs and knowledge graphs to embed canonical medical guidelines and research into the diagnostic process, thereby contextualizing raw patient data with established medical science.

[0051] The invention provides the effect of generating a Quantifiable and Objective Diagnostic Metric, translating a multitude of complex inputs into a single, transparently calculated score that represents the likelihood of a disease.

[0052] An additional advantage is AI-Driven Prognostic Forecasting, where the system analyzes the diagnostic metric over time to predict disease trajectories, extending its utility from a single diagnostic event to long-term patient management.

[0053] A key effect is Enhanced Explainability and Clinician Trust, achieved through a dual-layer explainability framework that provides both technical attribution and a plain-language, clinically validated rationale, with automated safeguards to prevent AI-generated misinformation.

[0054] The invention achieves Regulatory Compliance by Design, being architected as a SaMD with an integrated compliance module that automates adherence to medical device regulations (e.g., ISO 14971, IEC 62304) and data privacy laws.

[0055] Finally, the a Closed-Loop Learning Mechanism provides the advantageous effect of continuous and safe system improvement, allowing AI models to be refined based on verified clinical outcomes under strict validation protocols..Fig.1

[0056] is a schematic diagram illustrating the high-level architecture of the ASSURED-AID system, showing its major functional modules, the core knowledge and AI components, and the flow of data between them.Fig.2

[0057] is a workflow diagram depicting an exemplary embodiment of the invention applied to tuberculosis (TB) diagnosis, and includes a conceptual representation of the AI prompt workflow for generating the TB Diagnostic Metric.Fig.3

[0058] is a schematic diagram illustrating the hardware and network architecture for the generation and integration of the ASSURED-AID diagnostic metric, depicting the internal core logic and its secure interoperability endpoint connecting to the external healthcare ecosystem.

[0059] The following is a full description of a preferred embodiment of the invention. The embodiments are described in such a way that the disclosure is clearly communicated. The level of detail provided, on the other hand, is not meant to limit the expected variations of embodiments; rather, it is designed to include all modifications, equivalents, and alternatives that come within the spirit and scope of the current disclosure as defined by the attached claims. Unless the context indicates otherwise, the term “comprise”; and variants such as “comprises”; and “comprising” throughout the specification are to be read in an open, inclusive meaning, that is, "including, but not limited to"; When "embodiment" or "an embodiment" is used in this specification, it signifies that a particular feature, structure, or characteristic described in conjunction with the embodiment is present in at least one embodiment. As a result, the expressions “one embodiment” and “in an embodiment"; that appear throughout this specification do not necessarily refer to the same embodiment. Furthermore, in one or more embodiments, the specific features, structures, or qualities may be combined in any way that is appropriate. Unless the content clearly demands otherwise, the singular terms “a”, “an”, and “the”; include plural referents in this specification and the appended claims. Unless the content explicitly mandates differently, the term “or” is normally used in its broad definition, which includes "and / or".

[0060] The use of any and all examples, or exemplary language (e.g. “such as”) provided with respect to certain embodiments herein is intended merely to better illuminate the invention and does not pose a limitation on the scope of the invention otherwise claimed. No language in the specification should be construed as indicating any non-claimed element essential to the practice of the invention.

[0061] The headings and abstract of the invention provided herein are for convenience only and do not interpret the scope or meaning of the embodiments. All publications herein are incorporated by reference to the same extent as if each individual publication or patent application were specifically and individually indicated to be incorporated by reference. Where a definition or use of a term in an incorporated reference is inconsistent or contrary to the definition of that term provided herein, the definition of that term provided herein applies and the definition of that term in the reference does not apply.

[0062] Groupings of alternative elements or embodiments of the invention disclosed herein are not to be construed as limitations. Each group member can be referred to and claimed individually or in any combination with other members of the group or other elements found herein. One or more members of a group can be included in, or deleted from, a group for reasons of convenience and / or patentability. When any such inclusion or deletion occurs, the specification is herein deemed to contain the group as modified thus fulfilling the written description that follows, and the embodiments described herein, is provided by way of illustration of an example or examples, of particular embodiments of the principles and aspects of the present disclosure. These examples are provided for the purposes of explanation, and not of limitation, of those principles and of the disclosure.

[0063] It should also be appreciated that the present invention can be implemented in numerous ways, including as a system, a method or a device. In this specification, these implementations, or any other form that the invention may take, may be referred to as processes. In general, the order of the steps of the disclosed processes may be altered within the scope of the invention. Various terms as used herein are shown below. To the extent a term used in a claim is not defined below, it should be given the broadest definition persons in the pertinent art have given that term as reflected in printed publications and issued patents at the time of filing.

[0064] As illustrated in, the ASSURED-AID system is a comprehensive, unified platform organized into several interoperable modules. The system operates by synergistically combining and processing two distinct but interrelated streams of information: Canonical Knowledge, which represents the corpus of objective, established medical science, and Patient-Specific Knowledge, which comprises the dynamic, multi-modal data unique to an individual patient.

[0065] Recommendation Module (110): This module functions as the intelligent "front door" to the diagnostic process. Its purpose is to move beyond static, one-size-fits-all testing panels and to recommend an optimal, personalized diagnostic pathway for each patient. It analyzes the initial clinical picture and recommends the most diagnostically valuable and resource-efficient next steps, such as which specific biomarkers to test or which imaging modality to employ. The Recommendation Module receives initial patient data, which can be manually entered symptoms, information from an EHR, or patient monitoring data from a connected device. This triggers a decision engine that is not a simple checklist but an AI model. In one embodiment, this model is a trained machine learning classifier, such as a Gradient Boosting Machine (e.g., using the LightGBM library) or a Bayesian Network, which has been trained on vast historical datasets of diagnostic pathways and their associated outcomes. This model queries a rich internal knowledge base, which in a preferred embodiment is a graph database (e.g., Neo4j AuraDB), containing structured canonical information. This includes: (1) digitized clinical practice guidelines (e.g., from NCCN or ASCO), where nodes represent clinical states and edges represent recommended actions; and (2) performance characteristics of available tests, including sensitivity, specificity, cost benchmarks, and typical turnaround times. The engine calculates the pre-test probabilities of various differential diagnoses and then simulates the potential diagnostic yield of available tests to recommend the one with the highest expected value in terms of diagnostic certainty per unit of cost and risk. This functionality introduces personalization at the very beginning of the diagnostic workflow. It optimizes the use of healthcare resources, reduces patient burden by minimizing unnecessary or invasive tests, and accelerates the time to a definitive diagnosis by guiding the clinician along the most probable diagnostic path.

[0066] Interpretation Module (120): This is the analytical core of the system. It is responsible for ingesting all available patient data, fusing it with deep medical knowledge, computing the single, quantifiable composite diagnostic metric, and generating a transparent, dual-layer explanation of the result. Its operation is executed through a sequence of advanced sub-modules:Input Sub-Module (121): This sub-module acts as a universal data adapter and pre-processor. It ingests heterogeneous data through standardized interfaces. For instance, it receives laboratory results via HL7 v2 messages, parses them to extract biomarker names and values, and standardizes units (e.g., converting all creatinine values to mg / dL). It ingests imaging studies by retrieving them from a PACS via DICOM protocols and can run a pre-processing pipeline (e.g., using the pydicom library) to extract metadata and normalize image properties. For unstructured clinical notes, it employs Natural Language Processing (NLP) models (e.g., a fine-tuned version of BioBERT or a general-purpose library like spaCy with medical ontologies) to extract named entities (like diseases, symptoms, and medications) and their relationships.Processing Sub-Module (Federated Synthesis) (122): This sub-module performs the system's most novel function: the real-time fusion of patient data with canonical knowledge. The knowledge extractor component is an offline process that uses a powerful LLM (e.g., Google Gemini 2.5 Pro) to read, understand, and structure information from medical textbooks, peer-reviewed articles, and clinical guidelines. It converts this unstructured text into structured "triples" (Subject-Predicate-Object) that are used to build and continuously update a persistent Canonical Knowledge Graph. This graph, which may be hosted on a managed graph database service like Neo4j AuraDB or Amazon Neptune, models the universe of medical facts. When a patient's real-time data is received, the system constructs a transient Patient-Specific Knowledge Sub-Graph representing their current state. A federated synthesis then occurs. A complex query, constructed on-the-fly (e.g., a Cypher query for Neo4j), traverses nodes and relationships across both the patient's sub-graph and the canonical graph. For example, the query can start at the patient node, traverse to their lab result for an IL-6 value of 150 pg / mL, and then connect to the canonical graph to find that this value is > the a CRS_Risk_Threshold node, which in turn is DEFINED_BY a specific NCCN_Guideline node. This provides profound, machine-readable context.Metric Generation Engine (123): This engine takes the enriched, contextualized data from the federated synthesis as its input. It uses a sophisticated machine learning model to compute the final, unified composite diagnostic metric. In a preferred embodiment for complex, non-linear conditions, this is a Gradient Boosting Machine (e.g., using the XGBoost library) or a custom-designed neural network built with frameworks like TensorFlow or PyTorch. The model is trained to weigh the importance of different biomarkers dynamically based on the context provided by the knowledge graph. The output is a standardized score (e.g., 0.85 on a 0-1 scale) representing the aggregated probability of the condition.Explanation Framework (124): To ensure the system is not an opaque "black box," this framework provides a dual-layer explanation. The first layer, technical attribution, is generated using established XAI techniques implemented in libraries like SHAP or LIME. This produces a precise, machine-readable breakdown of which features (e.g., "IL-6 value," "Presence of fever") contributed most significantly to the final score. The second layer, the plain-language rationale, is generated by an LLM (e.g., Gemini 2.5 Pro) in a secure, sandboxed environment. This LLM receives the technical attribution scores and a prompt to explain their clinical significance. Crucially, before being displayed, the factual claims within the generated text are programmatically fact-checked by converting them back into queries against the canonical knowledge graph to mitigate the risk of LLM hallucination and ensure the explanation is clinically safe and accurate. This modular, multi-stage process for interpretation bridges the gap between purely statistical AI and knowledge-based expert systems, creating a tool that reasons much like a master clinician: it considers all available data, dynamically weighs it against a vast body of accumulated knowledge, and can explain both the quantitative and qualitative basis of its conclusion.

[0067] Prognosis Module (140): This module provides a forward-looking view by forecasting the patient's likely health trajectory, extending the system's utility from a single diagnostic event to longitudinal patient management. The module takes the composite diagnostic metric, particularly its trend over time, as input. It utilizes specific statistical and machine learning models suited for time-series forecasting. In one embodiment, it may use an ARIMA (Autoregressive Integrated Moving Average) model, implemented using the Python statsmodels library, to project the metric's future path. For more complex predictions involving time-to-event outcomes (e.g., predicting progression-free survival), it can employ a Cox proportional-hazards model using a library like lifelines. For capturing highly complex, non-linear temporal patterns, a Recurrent Neural Network (RNN), specifically an LSTM (Long Short-Term Memory) network built in TensorFlow / Keras, can be used. It transforms a static diagnostic score into a dynamic prognostic and monitoring tool, enabling clinicians to be more proactive in their management, anticipate complications, and provide patients with more informed expectations.

[0068] Compliance Module (130): As an internal governance and quality assurance layer, this module underpins the entire system's operation, ensuring its trustworthiness, safety, and regulatory adherence as a medical device. It follows the "compliance-by-design" philosophy, building safety, security, and auditability into the core architecture, rather than treating compliance as a post-hoc documentation exercise. This is a deeply integrated suite of services:Design Control and Architectural Enforcement: It maintains a software bill of materials (SBOM) and tracks the version and validation status of every software module and AI model, enforcing the modularity required by IEC 62304.AI-Based Risk Management: It implements a risk management file according to ISO 14971, with specific mitigations for AI risks. This includes runtime checks to ensure AI outputs adhere to pre-defined standards (e.g., Supported by Source Grounding in the explanation's rationale). This logic runs continuously, acting as an automated quality check on the AI's reasoning.Audit Logging and Traceability: Every critical action, from the data used to generate a specific metric to a clinician's interaction with the user interface, is recorded as an immutable, cryptographically signed entry in a dedicated audit log. In a preferred embodiment, this would be implemented using a service like Amazon Quantum Ledger Database (QLDB) to ensure a tamper-proof, verifiable history for regulatory audits.Data Privacy and Security: The module enforces security policies at every layer. Data in transit is protected with TLS 1.3, and data at rest is encrypted using AES-256. For processing highly sensitive data, the AI models may be deployed within a confidential computing environment (e.g., AWS Nitro Enclaves), ensuring the data remains encrypted even during processing. It manages user authentication and authorization, ensuring compliance with DPDP Act, 2023, GDPR, and HIPAA.QMS Integration: It provides APIs to interface seamlessly with commercial Quality Management Systems like Veeva Vault or Greenlight Guru, allowing for automated documentation of design changes, validation reports, and risk assessments.

[0069] Learning Loop (150) (Continuous Improvement): This module grants the system the ability to learn from real-world experience, ensuring its long-term accuracy and preventing model degradation over time. The process is a governed form of Reinforcement Learning from Human Feedback (RLHF). When a verified clinical outcome becomes available (e.g., a pathology report confirms a diagnosis, or a patient achieves remission), this "ground truth" is captured. This outcome is used to generate a "reward signal" for the system's previous prediction. A model retraining pipeline, which can be orchestrated by an MLOps platform like Kubeflow or MLflow, uses these reward signals to fine-tune the weights of the AI models in the Recommendation and Interpretation modules. The entire process is gated by the Compliance Module (130). Any newly retrained model is first deployed in a "shadow mode" for a validation period. Only after its performance is verified to be equal or superior to the current deployed version, and after this validation is formally documented, is the new model approved for promotion into the live clinical environment. It provides a safe and controlled mechanism for an AI medical device to continuously improve in the field. This overcomes a major hurdle for AI in regulated industries, creating a system that adapts to new medical knowledge and evolving disease patterns without compromising patient safety.

[0070] Exemplary Embodiment: TB-ASSURED-AID (Infectious Disease Case): To illustrate the invention's tangible application in a high-impact, resource-constrained setting, a specific embodiment for Tuberculosis (TB) diagnosis, termed TB-ASSURED-AID, is described in detail with reference to. This embodiment showcases how the modular architecture of the ASSURED-AID platform is specifically configured to address the complex challenges of TB screening and diagnosis, in alignment with national guidelines such as India's National Tuberculosis Elimination Program (NTEP) and the WHO-ASSURED criteria.

[0071] The TB-ASSURED-AID system is designed for deployment in community-based Active Case Finding (ACF) programs, often executed by frontline health workers. A health worker equipped with a standard tablet or mobile phone can initiate the diagnostic workflow. The Recommendation Module (210) first guides the worker through a simple questionnaire covering symptoms and risk factors (e.g., history of household contact). Based on this initial data, the module provides an instant recommendation, such as "Refer for Sputum Smear and NAAT" or "Schedule Follow-up." This workflow is fundamentally User-friendly, Rapid, Equipment-free (at the point-of-care), and Deliverable, making it ideal for field use. By optimizing the allocation of more expensive tests like NAAT, it also ensures the process is Affordable.

[0072] Following test recommendations, the Interpretation Module (220) receives various TB-specific inputs to generate the comprehensive TB Diagnostic Metric (TBDM). These inputs are diverse, reflecting the multi-modal nature of TB diagnosis, and may include, but are not limited to: a Symptom Score (SS) from the initial screening; a Chest X-Ray Score (CXR Score), often generated by a subordinate AI model trained to detect TB-indicative patterns; a Smear Microscopy Score (SMS); a NAAT Score (NATS) from a test such as Xpert MTB / RIF; and a Lateral Flow Assay Score (LFAS) from a urine LF-LAM test. It also includes canonical knowledge about TB diagnosis and treatment.

[0073] The first step in generating the TBDM within the Interpretation Module is Feature Quantization and Normalization. Each raw biomarker result, B_i, is programmatically converted into a standardized, normalized score, S_i, on a continuous scale of 0 to 1, where 0 represents no evidence of TB and 1 represents the strongest possible evidence from that single test. This is achieved via predefined mapping functions, f_i, which can be look-up tables or simple algorithms. For example, S_sms = f_sms(Smear_Result) might map categorical results to numerical scores like {Negative: 0.0, Scanty: 0.2, 1+: 0.5, 2+: 0.7, 3+: 0.9}. Similarly, S_nats = f_nats(NAAT_Result) would map {Negative: 0.0, Positive: 0.95}, reflecting the high specificity of the test.

[0074] The second and most novel step is the LLM-Mediated Contextual Weight Generation. Instead of relying on a fixed set of weights for each biomarker, the system's LLM (e.g., a secured instance of Google Gemini 1.5 Pro) dynamically generates a set of contextual weights, W_i, based on the complete patient profile. The LLM is provided with a structured prompt, for example, a JSON object, containing the patient's context (e.g., age, comorbidities, HIV status) and is tasked with reasoning about the relative diagnostic importance of each available test for that specific patient. For instance, if the prompt indicates the patient is HIV-positive, the LLM, having been trained on or having access to WHO guidelines via a RAG-on-Graph mechanism, would increase the diagnostic weight w_lfas for the LF-LAM test and potentially decrease the weight w_sms for smear microscopy, as smear sensitivity is known to be lower in this population. The LLM's output is a structured object containing not only the dynamically calculated weights {w_ss, w_cxs, ...} but also a "reasoning" string explaining why it chose those weights, which can be logged for audit purposes.

[0075] The third step is the Mathematical Computation of the TBDM (230). The final TBDM is computed as a contextually weighted average of the normalized scores from Step 1, using the dynamic weights from Step 2. The mathematical formula is a straightforward summation: TBDM = Σ(W_i * S_i) for all available biomarkers i. For a specific HIV-positive patient, if the LLM assigned weights {w_sms: 0.10, w_nats: 0.45, w_lfas: 0.30, w_ss: 0.15} and the normalized scores were {S_sms: 0.2, S_nats: 0.95, S_lfas: 0.8, S_ss: 0.7}, the TBDM would be calculated as: TBDM = (0.10 * 0.2) + (0.45 * 0.95) + (0.30 * 0.8) + (0.15 * 0.7) = 0.7925. A person of ordinary skill in the art would understand that this process, prompting an LLM API with structured patient data to receive a JSON object of weights, then applying those weights in a simple summation, is a concrete and implementable method. This is a significant inventive leap over static systems, as it embeds complex clinical reasoning directly into the metric's calculation.

[0076] Finally, the Interpretation Module provides a clear Output and integrates with the Clinical Workflow. The health worker's device displays the TBDM (79%) with a simple, color-coded risk level (e.g., "High Risk"). This is accompanied by a plain-language rationale generated by the system's Explanation Framework: e.g., "High TB Likelihood. The positive NAAT and LAM tests strongly suggest active TB, which is common in HIV-positive patients." The system then provides an explicit next-step instruction, such as "Refer patient to the nearest linked DOTS center immediately." To close the loop, the system can be configured to automatically transmit this result and referral information to the national health platform (e.g., Nikshay in India), ensuring seamless data flow and patient tracking.

[0077] This end-to-end workflow for TB-ASSURED-AID makes it highly Sensitive and Specific by intelligently combining multiple diagnostics and weighting them according to the patient's specific clinical context. By running on standard hardware and providing simple, actionable guidance, the system fulfills all principles of the WHO-ASSURED criteria, enabling the deployment of state-of-the-art diagnostic reasoning to the front lines of public health.

[0078] Hardware and Integration Aspects (): The ASSURED-AID system is architected to be deployed in various hardware configurations and integrate seamlessly with existing healthcare IT infrastructure.illustrates the architecture, emphasizing the secure generation of the diagnostic metric within a core system and its controlled interaction with the outside world. The architecture deliberately decouples the complex, sensitive AI core from the array of external devices. The AI SaMD Server (350) represents the secure computing environment where the diagnostic metric is generated. This core contains the Knowledge Resources (e.g., knowledge graphs), the AI / ML Pipeline (e.g., LLM servers, metric models), and an overarching Security & Compliance Layer.

[0079] The system does not interface directly with diagnostic hardware. Instead, an X-ray Unit (305) transmits imaging data to a Picture Archiving and Communication System (PACS) (315), while a Lab Analyzer (310) sends results to a Laboratory Information System (LIS) (320). The system’s data ingestion layer then interfaces with these established hospital systems. Specifically, the DICOM Gateway (350) is responsible for retrieving and processing imaging data from the PACS (315), and the HL7 Engine (345) is responsible for parsing and structuring clinical data from the LIS (320).

[0080] Core Processing Pipeline and AI Components: Once ingested, data enters a sophisticated processing pipeline. The DICOM Gateway (350) and HL7 Engine (345) feed structured data into the Recommendation Module (360), which performs initial analysis or flags key findings. These recommendations are then passed to the Interpretation Module (365), a central component that synthesizes information to generate a primary diagnostic interpretation. For deeper contextual understanding and the generation of natural language explanations, the Interpretation Module (365) queries the LLM Server (355). The LLM Server (355) leverages a comprehensive Knowledge Graph DB (375), which stores structured medical ontologies and relationships. This knowledge base is kept current and can be expanded via the Federated KG Connector (380). The output of the Interpretation Module (365) can feed into a Prognosis Module (370) to generate predictions about future outcomes.

[0081] Integration, User Interaction, and Compliance: The system's primary external-facing interface is the FHIR API (340). This standardized API serves multiple critical functions: it allows the Clinician App (330) to securely submit queries and receive results; it enables bidirectional communication with the EHR / HIS (325) for retrieving patient history and writing back final reports from the Interpretation (365) and Prognosis (370) modules; and it facilitates knowledge acquisition for the Federated KG Connector (380). Throughout the entire process, the Compliance Module (335) provides overarching governance. As shown by its connections to the data gateways and all analytical modules [(350), (345), (360), (365), (370)], it is responsible for auditing, logging, and enforcing security and data privacy policies, ensuring the entire workflow is secure, auditable, and compliant with healthcare regulations.

[0082] This architecture ensures that all data flows through a controlled, logged, and secure point, which is monitored by the Compliance Module. It supports both cloud and edge deployments and allows for continuous learning (model retraining) to occur on a dedicated high-power server (either in the cloud or on-premises) without disrupting the live diagnostic service. This design is fundamentally secure, scalable, and readily integrable into modern healthcare environments.

[0083] The present invention produces a significant technical effect by transforming the conventional, fragmented diagnostic workflow into a unified, adaptive, and technically robust process. It solves the technical problem of integrating heterogeneous, multi-modal data by employing a federated synthesis engine that fuses patient-specific data with a canonical knowledge graph, thereby generating a contextually enriched data representation that is beyond mere statistical aggregation. The system's generation of a single, quantifiable diagnostic metric provides a standardized and objective technical output, while the integrated verification component for the AI-generated explanation provides a crucial technical safeguard against misinformation, a known failure mode of large language models. This closed-loop learning architecture ensures the system's technical parameters dynamically improve over time, enhancing its long-term accuracy and robustness.

[0084]

[0085] The industrial applicability of this invention is extensive, as it is designed to be manufactured, deployed, and sold as a regulated Software as a Medical Device (SaMD) within the global healthcare and medical diagnostics industry. Its disease-agnostic and modular architecture allows for its application across numerous clinical specialties, including oncology, cardiology, and infectious diseases, making it a versatile platform for medical device companies and health-tech providers. By adhering to WHO-ASSURED criteria and integrating seamlessly with existing hospital IT infrastructure (EHRs, LIS, PACS), the system is commercially viable for diverse markets, from advanced tertiary hospitals to resource-limited, point-of-care settings, offering a valuable tool to improve diagnostic accuracy, reduce healthcare costs, and enhance patient outcomes.

[0086] Citation List follows:

[0087] Sun, Yuan, et al. “Method and System for Diagnosing Diseases Using Multi-Omics Data.” World Intellectual Property Organization, WO2024192175A1, 2024.

[0088] Sengupta, Arjun, et al. “Systems and Methods for Biomarker-based Disease Diagnosis.” United States Patent and Trademark Office, US20220292674A1, 2022.

[0089] Liu, Yong, et al. “Artificial Intelligence Based Multimodal Diagnostic System.” United States Patent and Trademark Office, US20240386015A1, 2024.

[0090] Zhou, Zhen, et al. “Biomarker Analysis Method and Device.” China National Intellectual Property Administration, CN114710707A, 2021.

[0091] Chen, Xuefeng, et al. “Device and System for Automated Biomarker Selection.” China National Intellectual Property Administration, CN111210366A, 2020.

[0092] Wang, Wei, et al. “Intelligent Medical Data Processing Method and Apparatus.” China National Intellectual Property Administration, CN109939901A, 2019.

[0093] Xu, Jian, et al. “Clinical Decision Support System Based on AI.” China National Intellectual Property Administration, CN110925640A, 2020.

[0094] Wang, Jun, et al. “Automated Analysis Platform for Clinical Biomarkers.” China National Intellectual Property Administration, CN113520787A, 2021.

[0095] Hodge, William, et al. “Diagnostic Platform and Methods for Biomarker Evaluation.” Canadian Intellectual Property Office, CA2650872C, 2010.

[0096] Deshpande, Sameer, et al. “System and Method for Disease Diagnosis Using Biomarker Panels.” United States Patent and Trademark Office, US20180068083A1, 2018.

[0097] Schultze, Joachim L., et al. “Systems and Methods for Medical Diagnostic Decision Support.” United States Patent and Trademark Office, US20040107124A1, 2004.

[0098] Vogl, Thomas, et al. “System for Integrating Biomarker Data in Medical Diagnostics.” United States Patent and Trademark Office, US11367184B2, 2022.

[0099] [NPL1] Popescu, C. P., et al. "Biomarkers in Critical Care: Application and Perspectives." Journal of Critical Care Medicine, vol. 7, no. 3, 2020, pp. 241-248.

[0100] [NPL2] Biomarkers Definitions Working Group. "Biomarkers and Surrogate Endpoints: Preferred Definitions and Conceptual Framework." NIH Biomarker - Bookshelf, 2024, pp. 1-3.

[0101] National Institutes of Health. “Biomarkers, EndpointS, and other Tools (BEST) Resource Fact Sheet.” 2024, pp. 2-5.

[0102] [NPL4] Smith, David, et al. "Quality Assurance in Laboratory Medicine: Practice Standards." Lab Medicine, vol. 38, no. 4, 2007, pp. 132-136.

[0103] [NPL5] Popescu, C. P., et al. "Biomarkers in Critical Care: Application and Perspectives." Journal of Critical Care Medicine, vol. 7, no. 3, 2021, pp. 241-248.

[0104] Binnicker, Matthew J., et al. "Direct Detection of Pathogens in Blood for Diagnosis of Sepsis." Journal of Clinical Microbiology, vol. 48, 2010, pp. 3236-3243.

[0105] [NPL7] Pai, Madhukar, et al. "Diagnostics for Tuberculosis: Global Policy Framework and Impact." BMJ Global Health, vol. 1, 2016, e000132.

[0106] [NPL8] World Health Organization. Diagnostics for Tuberculosis: Global Policy Framework. WHO, No. TDR / STI / IDE / 04.1, 2004.

[0107] Liu, Haoran, et al. "Emerging Diagnostic Tools for Infectious Diseases." Diagnostics, vol. 13, 2023, article 688.

[0108] United States Government Accountability Office. Artificial Intelligence in Health Care. GAO-22-104629, Sept. 2022.

Claims

A computer-implemented system for generating a quantifiable diagnostic metric and a corresponding prognostic forecast, the system implemented as a Software as a Medical Device (SaMD) and comprising:at least one computing device comprising one or more processors and memory configured to execute a plurality of modules, said modules comprising:a Recommendation Module (110) configured to process an initial clinical profile of a patient and, using an AI decision model, dynamically select and recommend a patient-specific subset of diagnostic tests from a plurality of available options to optimize a diagnostic pathway;an Interpretation Module (120) operatively coupled to the Recommendation Module, configured to:ingest and unify heterogeneous multi-modal data resulting from the recommended diagnostic tests;comprise a Federated Synthesis Engine (122) configured to access a canonical knowledge graph of medical information and perform a federated synthesis, thereby generating a contextually enriched patient data representation by fusing the patient's unified data with relevant information from the knowledge graph;comprise a Metric Generation Engine (123) configured to process said contextually enriched patient data representation using a machine learning algorithm to calculate a single, quantifiable composite diagnostic metric indicative of a disease state;an Explainability Sub-module (124) configured to generate a dual-layer explainable output for the composite diagnostic metric, the output comprising a technical attribution of input data contributions and a plain-language rationale, wherein said sub-module includes a verification component configured to programmatically cross-reference factual assertions within the plain-language rationale against the canonical knowledge graph to detect and mitigate AI-generated misinformation;a Prognosis Module (140) configured to analyze a temporal sequence of the composite diagnostic metric to forecast a disease progression trajectory; andan integrated Compliance Module (130) operatively connected to all other modules and configured to enforce a pre-defined risk management and data security framework, log system operations for auditability, and govern a closed-loop learning process.The system as claimed in claim 1, wherein the AI decision model of the Recommendation Module (110) is trained to prioritize diagnostic tests that adhere to WHO-ASSURED criteria, thereby optimizing the diagnostic pathway based on resource availability and the clinical setting.The system as claimed in claim 1, wherein the Federated Synthesis Engine (122) is configured to merge a transient, patient-specific data graph with the canonical knowledge graph to dynamically adjust the interpretive weighting of individual biomarkers based on the patient's broader clinical context, including comorbidities and demographic data.The system as claimed in claim 1, wherein the Compliance Module (130) further governs a closed-loop learning mechanism that updates the AI decision model or the machine learning algorithm using validated clinical outcomes, wherein said mechanism employs a federated learning architecture to enable model improvement across multiple sites without sharing raw patient data, and wherein updates are deployed only after passing a validation protocol in a sandboxed environment.The system as claimed in claim 1, wherein the Explainability Sub-module (124) is further configured to provide an interactive user interface element that, upon user interaction, displays the technical attribution, and is configured to flag and annotate any portion of the plain-language rationale that conflicts with verified facts in the canonical knowledge graph, thereby mitigating the risk of clinical error arising from AI-generated misinformation.The system as claimed in claim 1, further integrated with a healthcare IT infrastructure, wherein the system is configured to:receive said heterogeneous multi-modal data from Picture Archiving and Communication Systems (PACS) and Laboratory Information Systems (LIS) via secure DICOM and HL7 / FHIR protocols; andtransmit the composite diagnostic metric, the explainable output, and the prognostic forecast to an Electronic Health Record (EHR) system; thereby functioning as a technical bridge between disparate clinical information systems and the AI core.The system as claimed in claim 1, wherein the system architecture is disease-agnostic and extensible, configured to be adapted for a plurality of different medical conditions by modifying the canonical knowledge graph and the set of potential input biomarkers without redesigning the core modules.A computer-implemented method for generating a quantifiable diagnostic metric and a corresponding prognostic forecast using a Software as a Medical Device, the method executed by at least one processor and comprising the steps of:dynamically selecting, via a recommendation module, a patient-specific subset of diagnostic tests from a plurality of available options based on an initial clinical profile of a patient to optimize a diagnostic pathway;ingesting and unifying, via an interpretation module, heterogeneous multi-modal data from the selected tests and other clinical sources;performing a federated synthesis by fusing the unified patient data with relevant information from a canonical medical knowledge graph to generate a contextually enriched data representation;calculating, via a metric generation engine, a single quantifiable composite diagnostic metric from the contextually enriched data representation;generating and presenting a dual-layer explainable output, wherein the output includes the composite diagnostic metric, a technical attribution of input contributions, and a plain-language rationale, and wherein the method includes programmatically verifying the plain-language rationale against the medical knowledge graph to mitigate AI hallucination risks;forecasting, via a prognosis module, a disease progression trajectory by analyzing a temporal sequence of said composite diagnostic metric; andadaptively updating, under the governance of a compliance framework, one or more parameters of the AI models used in the selection or calculation steps based on validated clinical outcomes from a closed-loop feedback process.The method as claimed in claim 8, wherein the step of dynamically selecting further comprises an AI-driven decision algorithm prioritizing tests that adhere to WHO-ASSURED principles based on the patient's clinical setting.The method as claimed in claim 8, wherein the step of calculating the composite diagnostic metric employs a machine learning algorithm selected from the group consisting of a gradient boosting machine (GBM), a random forest, and a neural network, wherein the algorithm is trained to weigh inputs based on the context provided by the federated synthesis.The method as claimed in claim 8, wherein the step of adaptively updating employs a reinforcement learning framework where validated clinical outcomes serve as reward signals, and wherein model updates are validated in a shadow mode prior to clinical deployment to ensure compliance with medical device change control regulations.A non-transitory computer-readable storage medium having stored thereon program instructions that, when executed by one or more processors of a computing system, cause the system to perform the method of any one of claims 8 to 11.

Images