XUND

Medical assessment and clinical information retrieval are the core software products. Symptom and Illness Check combine a medical rule engine with a learned urgency classifier, while Health Check estimates future risks from structured personal information. Clinical Search uses natural language queries to retrieve and compose answers from clinical sources. These are distinct model driven functions rather than incidental productivity additions to an unrelated system.

APIs, content libraries, and patient navigation support their delivery, but the central value depends on interpreting symptoms, estimating risk, or finding relevant clinical knowledge. Medical professionals also review the underlying knowledge and assessment logic.

The pretriage model can raise urgency but is architecturally prevented from lowering the medical rule engine's assessment. This is a concrete restriction on automated decision behavior. Nevertheless, its model card describes sensitivity in terms of never missing critical cases, while the separate performance table reports 78.8% urgent case sensitivity. The current user manual expressly warns that urgent or serious conditions can be missed.

Those statements leave an unresolved inconsistency in the claimed reliability. Escalation-only logic reduces one route to unsafe behavior, but it does not reconcile an absolute completeness claim with the measured and acknowledged possibility of missed events.

The pretriage classifier uses one versus rest logistic regression with regularization and balanced class weights, combined with a separate medical rule engine. Health Check also uses logistic regression and identifies a large longitudinal biobank as its principal training source. The model cards name releases, input categories, outputs, and some validation choices.

Clinical Search instead describes answers grounded in medical repositories and local procedures, without identifying its underlying production language model. Complete training provenance, reproducible evaluation data, and release approval practices remain partly unspecified. Technical visibility is therefore stronger for the assessment models than for the newer search offering.

XUND identifies itself as developer of the assessment models and names infrastructure providers by purpose and processing location. These include T-Systems' sovereign Google Cloud service, MongoDB, Exoscale, Datadog, and the Smart Matters affiliate for relevant assessment products. Supplier review and data processing agreements are described as part of onboarding and continued oversight. This provides a substantial partial chain for the established services.

Clinical Search is not fully mapped in the same product register, and its underlying language model and provider specific training or retention arrangements remain unspecified. The portfolio therefore lacks a complete current model supply chain.

The June 3, 2026 pretriage model card identifies version 1.2.0 and an internal cross-validation method using annotated clinical cases. It reports 65.7% overall accuracy and 78.8% sensitivity for urgent cases, alongside stated layperson benchmarks. The card also acknowledges small datasets, uneven condition representation, and differences between expert summaries and patient questionnaires. These details make the evaluation more interpretable than an unsupported accuracy claim.

Independent confirmation in current patient deployments remains unspecified. Health Check's newer version heading and older version performance table describe different releases, so those results do not establish current version performance.

Clinical Search filters personal identifiers and requires users to avoid identifying patient information. Queries considered for offline model development undergo manual review, and only anonymous selections enter the training set. Queries not selected expire after one year; the retention period for selected training material remains unspecified. Assessment products separately use randomized check identifiers and minimize direct identifiers.

These are concrete data handling controls, but the complete retention and training boundary differs by product. Detailed technical measures for preventing fabricated clinical answers and investigating model safety incidents remain partly unclear, despite source links and medical content review.

Service privacy terms distinguish the application provider's controller responsibilities from XUND's processing role for Health Check information. A readable GDPR data processing agreement includes confidentiality, customer instructions, subcontractor obligations, and assistance with privacy rights. Its attached schedule specifically identifies Client Hub account and permission data, which limits conclusions about coverage of every clinical workflow.

The service privacy policy provides additional product specific handling detail. An accessible HIPAA BAA and complete contractual coverage across the portfolio remain unspecified. The European privacy arrangements provide substantive commitments without establishing that GDPR terms are interchangeable with a US business associate agreement.

The accessible ISO/IEC 27001:2022 certificate names XUND Solutions GmbH and Smart Matters Kft. Its scope covers the information security management system for medical software design, development, production, operation, and distribution. Certificate ISMS1531460 was issued on July 1, 2025, with stated validity through June 30, 2028. This is an actual current assurance artifact rather than only a badge or a request form.

The technical security material also describes access control, encryption, vulnerability management, and incident handling. The certificate establishes the stated management system scope; it does not independently validate the clinical accuracy of each model.

Symptom and Illness Check have an established European MDR medical device scope, with a current manual identifying XUND v1, its manufacturer, and basic UDI. The September 2026 instructions apply to product version 1.66.0 and later and set out intended use and restrictions. Clinical Search is different: its current product information says medical device approval is still underway.

Existing assessment certification therefore does not establish authorization for the newer search product's clinical answers. The portfolio contains both established regulated functions and a product with an incomplete approval process; no US FDA clearance is established by the European materials.

The model cards identify medical oversight, monitoring, and known population limitations. Health Check notes a training population concentrated in European ancestry and ages 40 to 69, along with healthy volunteer and geographic bias. The pretriage material describes medical review of features and ongoing subgroup and calibration surveillance. It also quantifies sensitivity for urgent cases, a clinically consequential subset.

These materials support informed review of model behavior and limitations. Comparative performance across demographic groups and current Clinical Search bias results remain unspecified, and an acknowledgment of underrepresentation is not a demonstrated measure of equivalent performance.

The current device manual provides a manufacturer support route for product problems and directs serious incidents to both XUND and the relevant authority. It also warns against relying on results when the software is malfunctioning. This creates a concrete path for reporting problematic behavior beyond general privacy rights or informal customer feedback. Separately, service materials describe availability and incident support commitments, but those concern service operation.

Output specific compensation, clinical indemnity, and a guaranteed correction remedy remain unspecified. The available recourse therefore supports reporting and investigation without promising reimbursement for an incorrect assessment or clinical answer.

Customers can integrate medical modules through an API, SDK, or web application and connect results with patient navigation inside their own service ecosystem. The product also supports clinical content integrations, including Thieme's eRef assistant, and links to provider networks or booking services. These are useful application and information interfaces. They do not establish a named production EHR connection, a specific clinical data exchange standard, or native chart writeback. The current integration evidence therefore explains embedding and workflow navigation more clearly than direct interoperability with a hospital or practice's system of record.

The medical modules run as hosted software in a containerized environment with shared infrastructure and logical separation between customers. German primary hosting is identified, but the location descriptions differ: the technical measures place backups in Switzerland, while the processor list identifies Austrian backup hosting and Swiss disaster recovery. The processor list also distinguishes services by product. Clinical Search lacks an equally complete processing map.

These details establish a European hosting architecture and tenant controls while leaving the precise backup arrangement and the full search inference path unresolved. An exclusive EU boundary would also conflict with the identified Swiss locations.

Licensing includes one region and its official language by default, with additional regions, languages, and optional services affecting the fee. Customers can select individual assessment modules and add medical content or navigation capabilities. This explains the product packaging and some cost drivers. Base license prices, usage allowances, implementation charges, and the cost of Clinical Search remain unspecified. The available information therefore supports understanding what may change a quote without establishing a representative purchase price for a particular deployment.

The September 2026 Symptom and Illness Check manual defines layperson use from age 16, with proxy assessment possible for adolescents aged 10 to 16 and people needing assistance. It excludes emergencies and known severe chronic disease. Health Check separately addresses adult preventive risk assessment, including cardiovascular, mental, tumor, musculoskeletal, and women's health topics.

Clinical Search serves healthcare professionals through clinical reference workflows, including Thieme's eRef assistant. These products cover different user groups and purposes. Their restrictions and validation scopes remain product specific, so professional search use does not broaden the assessment tools' permitted clinical use.