Introducing Clinical Conflict Detection Safeguards in Healthcare Agent Service
Generative AI is becoming increasingly prevalent in healthcare, and its significance is continuing to grow. Given the documentation-intensive nature of healthcare, generative AI presents an excellent opportunity to help alleviate this burden. However, to truly offset the clinician workload, it is crucial that content is checked for reliability and consistency before it is validated by a human. We are pleased to announce the private preview of our clinical conflict detection safeguard, available through our healthcare agent service. This safeguard helps users identify potential clinical conflicts within documentation content, regardless of whether it was generated by a human or AI. Identifying Clinical Conflicts: Seven Detected Categories Every conflict identified by the clinical conflict detection safeguard will indicate the conflict type and reference document content that constitutes the conflict so that the healthcare provider user can validate and take appropriate actions. Opposition conflicts: Normal vs abnormal findings of the same body structure E.g. Left breast: Unremarkable <> The left breast demonstrates persistent circumscribed masses. Negative vs positive statements about the same clinical entity E.g. No cardiopulmonary disease <> Bibasilar atelectasis Lab/vital sign interpretation vs condition E.g. Low blood sugar level at admission <> Patient was admitted with hyperglycemia Opposite disorders/symptoms E.g. Hypernatremia <> Hyponatremia Sex information opposites E.g. Female patient comes in with ... <> Testis: Unremarkable Anatomical conflicts: Absent vs present body structures E.g. Cholelithiasis <> The gallbladder is absent History of removal procedure vs. present body structure E.g. Bilat Mastectomy (2010) <> Left breast: solid mass Conducted imaging study versus clinical finding of body structure E.g. Procedure: Chest XR <> Brain lesion Laterality mismatch of same clinical finding E.g. Results: Stable ductal carcinoma of left breast. <> A&P: Stage 0 stable ductal carcinoma of right breast. Value Conflicts: Condition vs. lab / vital sign / measurement E.g. Hypoglycemia <> Blood Gluc 145 Conflicting lab measurement on same timestamp E.g. 02/11/2022 WBC-8.0 <> 02/11/2022 WBC-5.5 Contraindication conflicts: Medication/substance allergy vs. prescribed medication E.g. He is allergic to acetaminophen. <> Home medication include Tylenol, ... Comparison conflicts: Increased/decreased statements vs. opposite measurements E.g. Ultrasound shows a 3 cm lesion in the bladder wall, previously 4 cm, an increase in size. Descriptive conflict: Positive vs unlikely statements of same condition E.g. Lungs: Pleural effusion is unlikely <> Assessment: Pleural effusion Conflicting characteristics of same condition E.g. Results: Stable small pleural effusion <> Impression: Small pleural effusion Multiple versus Single statement of same condition E.g. Findings: 9 mm lesion of upper pole right kidney <> Assessment: Right renal lesions Metadata conflicts: Age information in provided metadata vs documentation E.g. Date of Birth = “04-08-1990” Date of Service=”11-25-2024" <> A 42-year-old female presents for evaluation of pneumonia. Sex information in provided metadata vs documentation * E.g. Date of Service=”11-25-2024" Sex= “female” <> Finding: Prostate is enlarged A closer look Consider the following radiology report snippet: Exam: CT of the abdomen and pelvis Clinical history: LLQ pain x 10 days, cholecystectomy 6 weeks ago Findings: - New calcified densities are seen in the nondistended gallbladder. - Heterogeneous enhancement of the liver with periportal edema. No suspicious hepatic masses are identified. Portal veins are patent. - Gastrointestinal Tract: No abnormal dilation or wall thickening. Diverticulosis. - Kidneys are normal in size. The patient comes in post cholecystectomy for a CT of abdomen/pelvis. We can create a simple request to the clinical conflict detection safeguards like this: { "input_document":{ "document_id": "1", "document_text": "Exam: CT of the abdomen and pelvis\nClinical history: LLQ pain x 10 days, cholecystectomy 6 weeks ago\nFindings:\n- New calcified densities are seen in the nondistended gallbladder.\n- Heterogeneous enhancement of the liver with periportal edema. No suspicious hepatic masses are identified. Portal veins are patent.\n- Gastrointestinal Tract: No abnormal dilation or wall thickening. Diverticulosis.\n- Kidneys are normal in size.", "document_metadata":{ "document_type":"CLINICAL_REPORT", "date_of_service": "2024-10-10", "locale": "en-us" } }, "patient_metadata":{ "date_of_birth": "1944-01-01", "date_of_admission": "2024-10-10", "biological_sex": "FEMALE", "patient_id": "3" }, "request_id": "1" } The request provides the metadata for document text to allow for potential metadata conflict detections. The clinical conflict detection safeguard considers the document text together with the metadata and returns the following response: { "inferences": [ { "type": "ANATOMICAL_CONFLICT", "confidence_score": 1, "output_token": { "offsets": [ { "document_id": "1", "begin": 73, "end": 88 } ] }, "reference_token": { "offsets": [ { "document_id": "1", "begin": 153, "end": 165 }, { "document_id": "1", "begin": 166, "end": 177 } ] } } ], "status": "SUCCESS", "model_version": "1" } The safeguard picks up an anatomical conflict in the document text and provides text references using the offsets that make up the clinical conflict. In this case, it picks up an anatomical conflict between “cholecystectomy” (which means a gallbladder removal) and the finding of “New calcified densities are seen in the nondistended gallbladder”. The new densities in the gallbladder conflict with the statement that the gallbladder was removed 6 weeks prior. In practice The clinical conflicts detected by the safeguard can be leveraged in various stages of any report generation solution to build trust in its clinical consistency. Imagine a report generation application calling the clinical conflict detection safeguards to highlight potential inconsistencies to the HCP end user — as illustrated below — for review before signing off on the report. There are multiple conflicts in the example above, but the highlight shows inconsistently generated documentation. The normal statement about the lungs contradicts “small nodules in the left lung” findings, so the “Lungs are unremarkable” statement should have been removed. How to use Apply for private preview by filling in the form here. Once approved, users must provision a healthcare agent service resource in their Azure subscription to use the clinical safeguards API. When creating the healthcare agent service, make sure to set the plan to “Agent (C1)”. * This clinical safeguard does not define criteria for determining or identifying biological sex. Sex mismatch is based on the information in the metadata and the medical note. Please remember that neither clinical conflict detection nor healthcare agent service are made available, designed, intended or licensed to be used (1) as a medical device, (2) in the diagnosis, cure, mitigation, monitoring, treatment or prevention of a disease, condition or illness or as a substitute for professional medical advice. The use of these products are subject to the Microsoft Product Terms and other licensing agreements and to the Medical Device Disclaimer and documentation available here.Empower clinicians with trusted content and GenAI: MSD Manuals added to healthcare agent service
Healthcare professionals today face the challenge of meeting patient demands, managing administrative tasks, and staying updated with ever-growing medical knowledge — all while working to deliver the best care possible. Generative AI-powered healthcare agents present a promising solution for helping with the unique complexities of medical data and patient care. By implementing rigorous controls and leveraging credible knowledge sources, healthcare agents can help clinicians make more accurate and relevant data-driven decisions, with the goal of improving health outcomes, efficiency, and safety. Today, we are excited to announce the integration of the MSD Manuals (known as Merck Manuals in the US and Canada), one of the world’s most widely used medical information resources, into the healthcare agent service in Microsoft Copilot Studio. The service enables healthcare organizations to develop their own conversational scenarios powered by generative AI, while leveraging healthcare-specific safeguards and pre-built healthcare intelligence from credible sources. By registering a Generative Answers on Built-in Sources plugin through the healthcare agent service orchestrator, healthcare organizations can develop agents that enable clinicians to ask questions in natural language and receive specific, evidence-based answers grounded on MSD Manuals and other trusted sources. All responses include source attribution, AI disclaimer, evidence snippets, and traceable links to the original articles, adhering to responsible AI principles. The MSD Manuals are a global medical information resource and have been a trusted source of medical information for over 125 years. They provide current, evidence-based content across all specialties for physicians and other health care professionals. The content is written, updated, and peer-reviewed by leading academic experts, undergoing a rigorous editorial process to enforce its accuracy and relevance. MSD Manuals are also available through healthcare agent service within Microsoft Dragon Copilot, allowing clinicians to access credible medical content within the context of the patient encounter without leaving their workflow. This integration marks a significant milestone in the journey toward more intelligent healthcare solutions. By bridging the gap between vast quantities of medical knowledge and the need for precise, patient-specific guidance, healthcare agent service in Microsoft Copilot Studio assists healthcare organizations and professionals in making evidence-based decisions effectively, responsibly and confidently. Learn more: Healthcare agent service - Healthcare agent service | Microsoft Learn Meet Microsoft Dragon Copilot: Your new AI assistant for clinical workflow MSD Manuals
