Symphony for Medical Coding

• Symphony for Medical Coding is an innovative framework that automates clinical coding beyond fixed label sets by consulting external guidelines.
• It employs a four-pillar architecture—evidence extraction, index navigation, tabular validation, and code reconciliation—to deduce accurate, auditable codes from free-text documentation.
• The system achieves robust performance across diverse datasets and coding systems by integrating retrieval, reasoning, and validation in an agentic workflow.

Symphony for Medical Coding denotes an agentic, guideline-grounded approach to automated clinical coding in which free-text clinical documentation is translated into standardized codes by reasoning over the clinical narrative with direct access to official coding resources, rather than by predicting from a fixed label subset alone. In its most explicit form, the system introduced in 2026 is designed to operate across coding systems, to avoid retraining when ontologies change, and to emit span-level evidence linking each predicted code to supporting text (Edin et al., 31 Mar 2026). The broader research program around the same motif treats coding as an orchestrated workflow that may begin with speech or document ingestion, proceed through evidence extraction, terminology lookup, code validation, and reconciliation, and end in auditable integration with clinical and billing infrastructure (Haque et al., 2018).

1. Definition and problem setting

Symphony for Medical Coding addresses medical coding as the process of translating free-text clinical documentation—such as progress notes, discharge summaries, emergency department notes, prescriptions, and operation notes—into standardized, machine-readable codes drawn from large clinical ontologies. The relevant coding systems include ICD-10-CM for diagnoses in the United States, ICD-10 in the UK NHS, ICD-10-PCS for inpatient procedures in the United States, CPT and HCPCS for ambulatory procedures and services, and related vocabularies such as SNOMED CT and OPCS-4 (Edin et al., 31 Mar 2026).

The task is intrinsically difficult for reasons repeatedly documented across the literature. The code spaces are large and heavy-tailed; coding rules change annually; guideline logic differs across inpatient and outpatient settings and across national variants; and clinical text is noisy, abbreviated, and often distributed across multiple documents. Earlier work on automated coding from free text therefore framed the problem as long-document, multi-label prediction under class imbalance and code association, while more recent systems emphasized retrieval, evidence grounding, and guideline application (Sun et al., 2021, Sun et al., 2021, Baksi et al., 2024).

A central distinction in Symphony is the rejection of closed-set assumptions. Traditional supervised systems predict over a fixed set of labels learned from annotated corpora, whereas Symphony is explicitly described as operating across any coding system by consulting the applicable manuals and guidelines at inference time (Edin et al., 31 Mar 2026). This design choice aligns it more closely with human coding practice and with adjacent agentic frameworks that implement the Alphabetical Index, the Tabular List, and official coding rules as first-class resources rather than latent regularities (Motzfeldt et al., 4 Sep 2025).

2. Research lineage and precursors

The “symphony” idea emerged from multiple lines of work that each addressed a different bottleneck in clinical coding. One precursor treated automatic documentation of ICD-10 code descriptions from far-field speech as a real-time seq2seq ASR problem. In that setting, ceiling- or wall-mounted microphones, spectrogram inputs, a convolutional plus pLSTM encoder, an attention decoder, and an unsupervised medical n-gram LLM were assembled into an end-to-end workflow for capture, transcription, code-description mapping, validation, and EHR integration (Haque et al., 2018). That work showed 87% accuracy and 85% BLEU on a far-field ICD-10 speech dataset, establishing an early orchestration-oriented view of coding automation (Haque et al., 2018).

A second lineage focused on automated coding from lengthy discharge summaries as multi-label classification. Multitask Balanced and Recalibrated Network and closely related Multitask Recalibrated Aggregation Network models used shared BiGRU encoders, recalibrated aggregation modules, label-aware attention, CCS auxiliary prediction, and focal or BCE-based multitask losses to address long notes, class imbalance, and code associations on MIMIC-III (Sun et al., 2021, Sun et al., 2021). These systems remained fundamentally supervised, but they clarified three persistent requirements that reappear in Symphony: long-context robustness, code-association modeling, and interpretable per-label attention.

A third lineage introduced retrieval- and evidence-centered generative assistants. MedCodER decomposed coding into extraction, retrieval, and re-ranking, using GPT-4 to identify diagnoses and supporting evidence, dense retrieval over ICD-10 ontology descriptors and UMLS descriptions, and listwise re-ranking with validity filtering; it reported micro-F1 0.60 on ICD-10 top-1 coding and diagnosis extraction micro-F1 0.83 on its benchmark (Baksi et al., 2024). LLM-codex used a two-stage architecture in which GPT-4 produced high-recall code and evidence proposals and an LSTM verifier learned from both gold document labels and LLM-generated sentence-level silver labels, reaching micro-F1 0.611 in a few-shot MIMIC-III top-50 setting and micro-F1 0.302 on a rare-code subset (Yang et al., 2023).

The most direct conceptual precursor is Code Like Humans, which mirrored the NHS Analyze–Locate–Assign–Verify workflow and operationalized the ICD Alphabetical Index, Tabular List, and official coding guidelines as the basis for a multi-agent, full-ICD-10 system (Motzfeldt et al., 4 Sep 2025). Symphony is explicitly described as extending that framework with stronger a priori knowledge, reasoning, search, and verification (Edin et al., 31 Mar 2026).

3. Core architecture

Symphony is organized around four pillars aligned with expert coding practice: Evidence Extraction, Index Navigation, Tabular Validation, and Code Reconciliation (Edin et al., 31 Mar 2026). The Evidence Extraction agent scans the clinical narrative for conditions and procedures that warrant coding, together with modifiers such as laterality, acuity, complications, staging, and context like present versus history versus ruled-out. The Index Navigation agent consults the alphabetic index to identify appropriate index terms and pointers into the code hierarchy. The Tabular Validation agent navigates the tabular list to select the most specific valid code while applying inclusion and exclusion notes, specificity constraints, and rules such as combination coding or “code first/use additional.” The Code Reconciliation agent then removes mutually exclusive outputs, enforces sequencing logic, and checks internal consistency across the full assigned code set (Edin et al., 31 Mar 2026).

The end-to-end flow begins with one or more encounter documents. Symphony can concatenate multiple note types for an encounter; MDACE, for example, concatenated ICU notes to mitigate omissions across note types. Preprocessing includes canonicalization, de-duplication of headers, section detection, and tokenization suitable for long notes. Intermediate outputs remain structured—candidate concepts, evidence spans, index terms, tabular paths, validation notes, and final rationales—which supports downstream auditing and interface design (Edin et al., 31 Mar 2026).

This architecture differs sharply from fixed-head classifiers. In MARN and MRAN, the shared encoder and task heads directly optimize code probabilities over predefined ICD and CCS vocabularies (Sun et al., 2021, Sun et al., 2021). In Symphony, by contrast, the external coding manuals determine the search space and validation path at run time. A plausible implication is that the representation burden shifts from memorizing label distributions toward coordinating retrieval, reasoning, and rule application.

Agentic workflow learning offers a complementary perspective. MedDCR treats workflow design itself as a search problem in which a Designer proposes plans, a Coder compiles them, a Reflector scores them, and a Memory Archive preserves successful designs (Zheng et al., 17 Nov 2025). Although Symphony does not describe this same closed-loop optimization procedure, both systems share the premise that coding performance depends materially on workflow decomposition rather than on a single monolithic generation step.

4. Evidence grounding, retrieval, and reasoning

Evidence-grounded output is one of Symphony’s defining properties. For each predicted code, the system returns span-level evidence linking the code to the clinical text, including the modifiers necessary for specificity. On MDACE, where professional coders annotated evidence spans, Symphony provided evidence for nearly all correct predictions, with evidence coverage of 98.9%; overlap metrics were reported as IoU $> 0$ of 73.7%, IoU $> 0.5$ of 42.5%, character-level F1 of 0.459, ROUGE-L F1 of 0.506, and average span IoU of 0.471 (Edin et al., 31 Mar 2026).

This emphasis on evidence responds directly to limitations identified in earlier LLM coding systems. MedCodER argued that traditional datasets do not contain linked, coder-usable evidence spans and therefore weaken interpretability and auditability; its pipeline explicitly carried diagnoses, evidence sentences, retrieval candidates, and ranked codes together into the final output (Baksi et al., 2024). LLM-codex likewise demonstrated that sentence-level evidence can be learned without human-annotated training evidence by using LLM-generated silver labels, and it achieved evidence F1 0.713 on MDACE Profee under a single-sentence evaluation criterion (Yang et al., 2023). Across these systems, evidence is not auxiliary decoration but the substrate for verification, ranking, and human review.

Retrieval in Symphony is structured by the official coding process. The system consults the alphabetic index, tabular lists, and official guidelines through tools, but the paper does not specify an internal ranking algorithm such as BM25 or dense cosine retrieval for the core product (Edin et al., 31 Mar 2026). By contrast, MedCodER explicitly used cosine similarity over embedded ICD descriptors and BM25 for evidence alignment, while Code Like Humans used embedding retrieval over Alphabetical Index terms and top-$k$ term retrieval per evidence span (Baksi et al., 2024, Motzfeldt et al., 4 Sep 2025). Taken together, these results suggest that ontology access and evidence retrieval are indispensable, but that the central design question is not merely which retriever is used; it is how retrieval is coupled to code validation and reconciliation.

A common misconception is that modern coding LLMs can simply generate codes from raw notes with acceptable reliability. Several papers argue against this. GPT-4 in LLM-codex overpredicted codes, yielding high recall but low precision (Yang et al., 2023). MedCodER’s ablations showed that performance dropped when extraction, retrieval, or re-ranking were used in isolation (Baksi et al., 2024). Symphony’s own design can be read as a direct response: reasoning is decomposed, guidelines are externalized, and validation is explicit rather than implicit (Edin et al., 31 Mar 2026).

5. Empirical evaluation and generalization

Symphony was evaluated on two public benchmarks and three real-world datasets spanning inpatient, outpatient, emergency, and subspecialty settings in the United States and the United Kingdom (Edin et al., 31 Mar 2026). The public datasets were ACI-BENCH and MDACE. The proprietary datasets were a United States emergency department corpus with 563,153 notes and 11,148 unique ICD-10-CM codes, a United States ambulatory corpus with 2,250,380 notes and 17,561 unique ICD-10-CM codes, and a UK NHS neurology corpus with 10,675 notes and 667 unique ICD-10 codes (Edin et al., 31 Mar 2026).

Under restricted code-system evaluation, Symphony achieved $R=0.81 \pm 0.01$, $P=0.68 \pm 0.01$, and $F1=0.74 \pm 0.01$ on ACI, and $R=0.66 \pm 0.02$, $P=0.59 \pm 0.02$, and $F1=0.62 \pm 0.02$ on MDACE (Edin et al., 31 Mar 2026). Under full-code-system evaluation, which the paper presents as more realistic for deployment, Symphony obtained $F1=64.1 \pm 1.7$ on ACI, $> 0.5$0 on MDACE, $> 0.5$1 on ambulatory notes, $> 0.5$2 on emergency department notes, and $> 0.5$3 on the UK neurology dataset, surpassing GPT and CLH baselines in each setting reported (Edin et al., 31 Mar 2026). Procedure coding results were also reported: ICD-10-PCS on MDACE reached $> 0.5$4, CPT on ambulatory notes $> 0.5$5, and CPT on emergency department notes $> 0.5$6 (Edin et al., 31 Mar 2026).

The generalization claim is notable because Symphony is described as requiring no retraining to adapt across ICD-10-CM, UK ICD-10, ICD-10-PCS, and CPT; adaptation is achieved by changing the external resources queried by the agents (Edin et al., 31 Mar 2026). This contrasts with earlier classifier families, whose strong MIMIC performance remained tied to fixed label spaces and specific coding schemes. For example, MARN reported on MIMIC-III-full $> 0.5$7 ICD-9-CM codes with Macro-F1 11.6, Micro-F1 58.4, $> 0.5$8 75.4, and $> 0.5$9 60.2, while its top-50 ICD setting reached Macro-F1 68.2 and Micro-F1 71.8 (Sun et al., 2021). Those results are strong within their benchmark regime, but they do not remove the retraining requirement when the ontology or institutional setting changes.

Evidence-grounded retrieval systems also provide useful comparators. MedCodER reported micro-F1 0.60 for ICD-10 top-1 coding on its benchmark (Baksi et al., 2024). Code Like Humans, evaluated on MDACE with the full ICD-10 space, reported micro-F1 0.42 and macro-F1 0.27 for its base configuration, and emphasized its advantage on rare diagnosis codes (Motzfeldt et al., 4 Sep 2025). Symphony’s gains over these systems therefore appear to arise not from abandoning their core insights, but from strengthening the same ingredients—guideline grounding, search, verification, and reconciliation—inside a more mature agentic architecture (Edin et al., 31 Mar 2026).

6. Operational use, limitations, and future directions

Symphony is presented as deployment-oriented. It is accessible via a production-grade API and as MCP plugins for multi-agent frameworks, and its structured outputs are designed for UI display, RCM workflows, CDI review, and auditing (Edin et al., 31 Mar 2026). Privacy and regulatory constraints remain central: the proprietary evaluations were conducted under HIPAA- and GDPR-protected conditions, and some third-party LLMs could not be tested on PHI datasets because of deployment restrictions (Edin et al., 31 Mar 2026). Earlier orchestration blueprints similarly recommended HL7 or FHIR interfaces, secure logging, clinician confirmation workflows, and encryption of PHI in transit and at rest (Haque et al., 2018).

Human oversight remains important. Symphony emphasizes high precision, but its authors explicitly note annotation variability, specificity disagreements, and ongoing ambiguity in nuanced documentation (Edin et al., 31 Mar 2026). In the UK dataset, some disagreements appear to arise because Symphony selected more specific codes than the reference annotations, suggesting that benchmark discordance may sometimes reflect label noise rather than erroneous reasoning (Edin et al., 31 Mar 2026). Related studies echo this concern from different angles: MedCodER highlights the value of coder-visible evidence spans (Baksi et al., 2024), LLM-codex documents GPT-4 overprediction and the need for verification (Yang et al., 2023), and the hospital-course summarization literature shows that overlap metrics such as ROUGE can mis-rank clinically incorrect outputs, making coder-oriented factual assessment indispensable (Bi et al., 2024).

Known limitations are stated directly. Symphony’s CPT support is described as beta; detailed prompts and per-component numeric ablations are not disclosed; and black-box LLM components can still make occasional reasoning errors (Edin et al., 31 Mar 2026). More broadly, the literature identifies unresolved issues in rare-code coverage, multilingual adaptation, richer procedure coding, better handling of multiple speakers or tonal languages for speech-based capture, synthetic data generation for long-tail labels, and stronger calibration or uncertainty-based deferral (Haque et al., 2018, Seedat et al., 2024, Vo et al., 18 Nov 2025).

Future work therefore converges on a common agenda: broader guideline coverage, deeper subspecialty adaptation, stronger long-tail support, better evidence modeling, and tighter integration with EHR and coding infrastructure. A plausible implication is that Symphony for Medical Coding is not merely a single system name but a direction of travel for the field: away from closed-set prediction and toward auditable, ontology-aware, workflow-centered coding automation grounded in the same resources and constraints that govern expert human coders (Edin et al., 31 Mar 2026, Zheng et al., 17 Nov 2025).