Posts tagged Healthcare AI.
On 18 June 2026 Midjourney unveiled a whole-body ultrasound scanner with reconstruction and segmentation handled by learned models. A technical look at the physics, the pipeline and the validation limits, separating the verifiable specifications from the launch claims.
Read moreA concrete architecture for an on-premise clinical Retrieval-Augmented Generation system anchored to FHIR and DICOM: ingest, retrieval, local biomedical LLM, answers with verifiable citations, and the MDR/AI Act/GDPR constraints read as design choices.
Read moreA medical device with machine learning that changes after certification is, for the regulator, a different device. MDR and the FDA answer with very different tools and timelines.
Read moreAn NVIDIA GB10 workstation with 128 GB of unified memory holds models up to 200B parameters under a desk. For health data the constraint is not capacity but memory bandwidth: 273 GB/s.
Read moreHealthcare demands AI systems that are verifiable, explainable, and compliant with rigorous clinical and regulatory standards.
Read moreA language model in contact with records and reports has to stay inside the perimeter: PII redaction upstream, immutable logging for the EU AI Act, risk classification under MDR. The architectural constraints that follow.
Read moreCDA R2, FHIR and openEHR do not sit on the same plane of a health system: they answer three different questions — archiving, exchange, keeping data over time. The roles, the contact points, and the friction that appears when a standard ends up in the wrong place.
Read moreRegulation (EU) 2025/327 on the European Health Data Space has been in force since 26 March 2025. A technical reading of primary use, secondary use, the Health Data Access Body and the application timeline.
Read moreThe Italian Ministry of Health decree of 7 September 2023 (Official Gazette no. 249 of 24 October 2023) introduces the Sogei-run Gateway and the regional Healthcare Data Ecosystem. Document formats, IHE profiles and the FSE 2.0 feeding model as of September 2024.
Read moreThe International Patient Summary (EN ISO 27269) defines what a patient summary contains; MyHealth@EU how it travels between Member States. Why keep the two layers apart, and why semantic mapping stays the fragile part.
Read moreThe regulatory line between an ESP32 + MQTT node carrying a MAX30102 sensor and a medical device under Regulation (EU) 2017/745. What changes between a telemonitoring demo and something you can put in a patient's hands.
Read morePlacing a deep-learning prediction into the data flow of a Total Laboratory Automation system imposes constraints on protocols, latency and traceability that come before the quality of the model.
Read moreMONAI Label moves the cost of medical annotation from manual drawing to correcting a proposal: a REST server, plugins for 3D Slicer and OHIF, and active learning driven by model uncertainty.
Read moreBioBERT, ClinicalBERT and PubMedBERT differ in corpus, vocabulary and domain. A technical reading of their pre-training choices and the measurable effect on clinical tasks.
Read moreItaly's FSE 2.0 plan, presented to the State-Regions Conference in the briefing of 2 March 2022, shifts the Electronic Health Record from a document-centric CDA R2 model to a data model based on HL7 FHIR and regional Health Data Ecosystems. The technical knots of the transition.
Read moreTechnical notes on the most widely deployed open source healthcare integration engine: the source-transformer-destination pipeline, E4X parsing of HL7 v2 over MLLP, and where the model holds or gives way under load.
Read moreTaking a classifier into clinical production means deciding where data crosses the hospital boundary. Three concrete constraints: FHIR interoperability, data residency, regulatory validation.
Read moreOn 21 April 2020 NVIDIA and King's College London presented MONAI, a PyTorch library that gathers in one place the transforms, networks and metrics that every medical imaging lab had until now rewritten on its own.
Read moreTechnical notes on applying machine learning to paediatric short bowel syndrome cohorts: the shape of longitudinal clinical data, target definition, and the constraint of small sample size.
Read moreThe state of convolutional networks for skin-lesion classification in mid-2019: where the public benchmarks hold, and where integration into a laboratory workflow remains the open problem.
Read moreFHIR Release 4 (4.0.0, December 2018) is the first version with Normative content: the REST API, XML/JSON formats, base datatypes and a few clinical resources leave Trial Use with a backward-compatibility commitment. A technical note on the boundary of the Normative set and what changes for those building clinical systems.
Read moreThe nnU-Net preprint (arXiv 1809.10486) argues that, for medical image segmentation, a well-configured standard U-Net beats architectural variants. Notes on the data fingerprint, the three configurations and the provisional Medical Segmentation Decathlon results.
Read moreBy 2018, models on clinical records and diagnostic images beat the published baselines, but the constraint that decides everything is the data: scarcity, quality, treatment regime, and GDPR.
Read moreFrom 25 May 2018 processing health data starts from a prohibition and ten exhaustive derogations. The Article 9(2) derogation you invoke decides who reads the data, where it may travel, and when a clinical dataset must stay separate from a research one.
Read moreControlled clinical terminologies are not interchangeable: SNOMED CT describes the clinical picture, LOINC names observations and reports, ICD classifies for statistics and reimbursement. Boundaries, licensing and Italian adoption as of 2017.
Read moreSleep stages are derived from physiological signals read according to codified rules. Scoring standards, data exchange formats, actigraphy and the EU regulatory path for sleep-analysis software as of early 2017.
Read moreAutomatic sleep staging between the historical actigraphy rules and time-series classifiers on polysomnographic signals, constrained by class imbalance and clinical validation.
Read moreA technical read of the Medical Imaging Interaction Toolkit at release 2016.03: the ITK + VTK stack, DataStorage as the shared data model, BlueBerry sitting on the CTK Plugin Framework and CppMicroServices, and the limits that remain on the regulatory side.
Read moreThe technical rules for Italy's Electronic Health Record set out a federated architecture built on signed CDA R2 documents and IHE XDS.b/XCA profiles, with INI as an interoperability gateway between regional records.
Read moreA stable Reference Model underneath, clinician-governed archetypes in ADL on top. How openEHR keeps the software model separate from clinical concepts, and what that costs and yields for interoperability and maintenance.
Read moreFHIR DSTU 1 (v0.0.82) exposes clinical entities as REST resources in XML or JSON. Reading notes on its design choices against HL7 v2 and CDA R2.
Read moreA technical reading of DICOM in 2013: the object model, SOP Classes, DIMSE services, transfer syntaxes, and the move under way towards the REST services WADO-RS, QIDO-RS and STOW-RS. What you build with the Open Source toolkits.
Read moreDecree-law 179/2012 gives Italy's Electronic Health Record a national statutory basis while deferring its operational content to a Prime Ministerial decree that has yet to be issued. What Article 12 establishes, how it rests on the Digital Administration Code, and which technical questions remain open.
Read moreHow an HL7 v2 message is built: pipe-delimited encoding, segments, message types, MLLP transport, and why the variability between implementations is the real cost of hospital integration.
Read moreOpenCDS exposes clinical logic as an HL7 DSS web service that takes a virtual Medical Record and returns structured recommendations, with Drools as the execution engine. A technical note on the architecture, the authoring bottleneck and the limits as of 2012.
Read moreNotes on the JAMIA paper describing cTAKES (Savova et al., 2010): its UIMA annotator architecture, the UMLS dictionary lookup, NegEx-based negation, and the component figures measured on the Mayo corpus.
Read moreHow the National Library of Medicine's MetaMap links English biomedical text to Concept Unique Identifiers in the UMLS Metathesaurus: the lexical-variant pipeline, and the practical limits around ambiguity, language and coverage.
Read moreOpenCV 1.0 is not a medical imaging library, yet its 2D operations cover much of what mammography, retinoscopy, dermatology and microscopy actually need. What the 1.x line offers in 2008, where its domain ends, and why DICOM and CT/MR volumes stay out.
Read moreDrools 4.0 (July 2007) as an execution engine for clinical logic: the Rete algorithm, a working memory of POJOs and the unresolved data-binding problem, set against Arden Syntax and GELLO.
Read moreArden Syntax 2.5 separates clinical logic from the chart's application code, yet leaves access to local data inside non-standardised curly braces. GELLO, approved by HL7 in 2005, tries to close that gap by tying expressions to the RIM.
Read moreAbove ITK and VTK there is a missing layer: software a clinician can actually open and use. 3D Slicer occupies that layer, and the NA-MIC grant of September 2004 has just changed how it is built.
Read moreThe state of WEKA in mid-2003: the ARFF format, the bundled classifiers (J48, NaiveBayes, SMO), and why in quantitative clinical research the validation protocol weighs more than the individual algorithm.
Read moreTechnical notes on the Insight Toolkit, the C++ library funded by the National Library of Medicine under the Visible Human Project: template architecture, dataflow pipeline, CMake build and BSD licence.
Read moreWithin a few years R has become a practical tool for analysing high-dimensional gene-expression data. Technical notes on why a free statistical language fits microarrays, and on what the Bioconductor project is trying to standardise.
Read more
