FHIR-Compliant Digital Twin Representations

• FHIR-compliant digital twin representations are machine-readable patient models structured using HL7 FHIR for personalized monitoring and decision support.
• The methodology employs sequential NLP modules—NER, concept normalization, relation extraction, and FHIR assembly—to transform unstructured clinical notes into structured data.
• Empirical results demonstrate superior precision, recall, and interoperability compared to baselines, highlighting its potential for advanced healthcare analytics.

A FHIR-compliant digital twin refers to a machine-readable, standards-based patient representation structured using the HL7 FHIR (Fast Healthcare Interoperability Resources) specification, with all clinical states, interventions, and temporal trajectories derived from source data such as unstructured clinical notes. These digital twins serve as interoperable virtual replicas of patients, supporting personalized monitoring, predictive modeling, and decision support in healthcare. Recent research presents detailed semantic NLP pipelines enabling the transformation of free-text EHR data into validated digital twin artifacts structurally and semantically conformant to FHIR (Brens et al., 9 Jan 2026).

1. Semantic Pipeline for Digital Twin Construction

The end-to-end pipeline operates over four sequential NLP modules, each responsible for a critical transformation from clinical prose to structured FHIR resources.

1. Named Entity Recognition (NER): Utilizing a ClinicalBERT model with a token-classification (BIO-tagging) head, the NER module identifies clinical span types—specifically Condition, Medication, Observation, and Temporal entities. Model input combines WordPiece embeddings, segment and positional embeddings, and concatenated character-level features derived from a CNN. Training data comprises synthetic discharge notes (100 patients, ~1,200 notes) generated from MIMIC-IV Demo structured tables, with a train/dev/test split and silver-standard annotations.
2. Concept Normalization: Each recognized span is mapped to standardized codes via a two-stage approach. Initial candidate codes—SNOMED-CT or ICD-10 for conditions, RxNorm for medications, LOINC/SNOMED for observations—are retrieved using UMLS dictionary matching and SapBERT embedding-based nearest neighbor search. Disambiguation leverages cosine similarity between the ClinicalBERT context vector and SapBERT code embeddings, with a threshold $\tau = 0.75$ for acceptance. Where available, both SCTID and ICD10 codes are recorded for maximal interoperability.
3. Relation Extraction: Relations captured include has-dosage (Medication→DosageInstruction), symptom-of and indicates (Observation→Condition), and treats (Medication→Condition). Modeling employs ClinicalBERT with explicit entity markers and a one-layer FFNN classifier, trained on i2b2/VA 2010 annotated relations and synthetic mappings from MIMIC-IV Demo. Decoding retains relations with classifier softmax probability ≥ 0.5.
4. FHIR Assembly & Validation: Extracted entities and relations are algorithmically assembled into FHIR-compliant bundles, representing the minimal profile: Patient, Condition, Observation, and MedicationRequest. Each resource is syntactically and semantically validated against the official FHIR schema.

2. FHIR Resource Profiles and Mapping Approach

Digital twin assembly targets a strictly defined minimal FHIR R4 profile:

Resource	Identifier/Reference Example	Code Systems Utilized
Patient	"id": "pt-123"	–
Condition	subject.reference → Patient/{id}	SNOMED-CT, ICD-10
Observation	subject.reference → Patient/{id}	LOINC, SNOMED-CT
MedicationRequest	subject.reference → Patient/{id}	RxNorm

• Patient: Resource comprises demographic fields (gender, birthDate) inherited from original EHR note.
• Condition: Contains semantic links to Patient, codified with one or both of SNOMED-CT and ICD-10 codes, and includes clinical/verification statuses.
• Observation: Encodes physiological measurements, lab results, or symptoms, mapping codes to LOINC or SNOMED, with quantitative results and timestamps from temporal parsing.
• MedicationRequest: Relates patient, medication (RxNorm code), and dosage, referencing structured or free-text equivalents via has-dosage relations.

3. Implementation Mechanics and Data Artifacts

The resource assembly process is programmatically formalized. Assembly pseudocode (Python-style) iterates over recognized entities, constructing each FHIR resource and collectively validating the output:

def assemble_bundle(patient_id, entities, relations): bundle = { "resourceType": "Bundle", "type": "collection", "entry": [] } bundle["entry"].append({"resource": make_patient_resource(patient_id)}) for e in entities["Condition"]: cond_res = make_condition(e, patient_id) bundle["entry"].append({"resource": cond_res}) for e in entities["Observation"]: obs_res = make_observation(e, patient_id) bundle["entry"].append({"resource": obs_res}) for e in entities["Medication"]: medreq = make_medication_request(e, relations, patient_id) bundle["entry"].append({"resource": medreq}) validated_entries = [] for entry in bundle["entry"]: if validate_fhir(entry["resource"]): validated_entries.append(entry) bundle["entry"] = validated_entries return bundle

A representative FHIR Bundle JSON snippet:

{ "resourceType": "Bundle", "type": "collection", "entry": [ { "resource": { "resourceType": "Patient", "id": "pt-123", "gender": "male", "birthDate": "1958-07-19" } }, { "resource": { "resourceType": "Condition", "id": "c1", "subject": { "reference": "Patient/pt-123" }, "code": { "coding": [ { "system":"http://snomed.info/sct","code":"38341003","display":"Hypertensive disorder" } ] }, "clinicalStatus": { "coding":[{"system":"http://terminology.hl7.org/CodeSystem/condition-clinical","code":"active"}] }, "verificationStatus": { "coding":[{"system":"http://terminology.hl7.org/CodeSystem/condition-ver-status","code":"confirmed"}] } } }, { "resource": { "resourceType": "Observation", "id": "o1", "subject":{"reference":"Patient/pt-123"}, "code":{ "coding":[{"system":"http://loinc.org","code":"85354-9","display":"Blood pressure panel"}] }, "valueQuantity": {"value":145,"unit":"mmHg"}, "effectiveDateTime":"2023-01-01T10:00:00Z" } } ] }

4. Extrinsic Evaluation and Comparative Performance

The pipeline evaluation leverages the MIMIC-IV Clinical Database Demo, validated against the MIMIC-IV-on-FHIR reference. Metrics include precision, recall, $F_1$ (entity-level, relation-level), semantic completeness, and interoperability score:

• Precision: $\frac{TP}{TP + FP}$
• Recall: $\frac{TP}{TP + FN}$
• F1: $$2 \times \frac{precision \times recall}{precision + recall}$$
• Semantic Completeness: $$\frac{\text{number of required FHIR fields correctly populated}}{\text{total required fields in reference FHIR}}$$
• Interoperability Score: Combined measure of structural and semantic similarity to reference, incorporating both field-level and terminology-level concordance.

Empirical results (synthetic notes from 100 patients) benchmark the presented semantic pipeline against two baselines:

Method	NER F1	RE F1	Completeness	Interop
Rule-Based	0.72	0.55	62%	0.61
Naive Mapping	0.68	0.50	58%	0.58
Ours	0.89	0.81	91%	0.88

Ablation studies demonstrate the significance of normalization (interop drops from 0.88 to 0.65 if omitted) and relation extraction (completeness drops from 91% to 68%). Performance is highest when ClinicalBERT rather than base BERT is used for NER (F1 0.89 vs 0.81) (Brens et al., 9 Jan 2026).

5. Limitations, Open Challenges, and Future Extensions

A principal limitation is the reliance on synthetic, English-language notes and references semi-automated FHIR mappings, constraining direct generalizability. The pipeline does not model longitudinal or multi-encounter patient trajectories, which restricts temporal analytics and dynamic digital twin updating. The current relation schema omits familial, procedural, and allergy histories. Multilingual extension and scaling to larger, real-world and cross-institutional datasets are recognized as essential next steps.

Possible directions include:

• Incorporation of laboratory and imaging reports as valid FHIR Observation types.
• Modeling temporal relations across encounters for longitudinal and dynamic digital twins.
• Integration with SMART on FHIR applications, enabling real-time clinical decision support.
• Expansion to capture richer biomedical relations (family history, procedures, allergies).

This approach—coupling transformer-based NER, ontology-grounded normalization, and relation extraction to a rigorously defined FHIR minimal profile, validated against the official specification—ensures semantic fidelity and syntactic interoperability for downstream analytics and exchange (Brens et al., 9 Jan 2026).