A Hybrid Rule‐Based and Large Language Model Artificial Intelligence System for Electrodiagnostic Reporting: Two‐Center Retrospective Evaluation

ABSTRACT

Introduction/Aims

Electrodiagnostic (EDX) studies comprise nerve conduction studies (NCS) and needle electromyography (EMG). However, EDX reporting is heterogeneous across laboratories and often requires time‐consuming documentation. We aimed to evaluate a hybrid rule‐based and constrained large language model (LLM) system that drafts EDX reports from structured EDX measurements.

Methods

We retrospectively analyzed EDX reports from two institutions. The system generated a structured Summary (standardized abnormal findings derived from EDX measurements using laboratory‐specific cutoffs), and then used a label‐constrained LLM to draft a diagnostic Interpretation (a clinician‐facing diagnostic conclusion) by selecting from a predefined diagnostic‐label inventory. We evaluated agreement with clinician diagnostic interpretation using a Coverage–Localization Index (CLI), which summarizes whether the same diagnostic labels were selected (coverage) and whether the site/level localization matched (localization).

Results

Among 500 screened examinations, 332 reports were included (241 primary‐care; 91 tertiary‐care). Overall CLI agreement between system and clinician diagnostic interpretation was high (mean total 96.7%, 95% CI 95.1–98.1), with lower agreement in more complex examinations.

Discussion

The rule‐based extractor captured all rule‐defined abnormal findings, and the constrained LLM drafted diagnostic interpretations that showed high agreement with clinicians, generating standardized draft reports. Only abnormal, in‐scope EDX reports were included, which limits direct inference about performance in an unselected clinical stream. Future work should validate the approach on consecutive routine EDX reports, including normal and diagnostically ambiguous studies, and quantify workflow impact such as editing time and clinician acceptability.

Abbreviations

AI

artificial intelligence

CB

conduction block

CC

chief complaint

CI

confidence interval

CLI

coverage–localization index

CMAP

compound motor action potential

CNS

central nervous system

CNUH

Chungnam National University Hospital

EDX

electrodiagnostic

EMG

electromyography

IRB

Institutional Review Board

LLM

large language model

MUAP

motor unit action potential

NCS

nerve conduction study

SNAP

sensory nerve action potential

TD

temporal dispersion

UI

user interface

1. Introduction

Electrodiagnostic (EDX) studies comprise nerve conduction studies (NCS) and needle electromyography (EMG). EDX interpretation is a learnable clinical skill with established training pathways; nonetheless, EDX reports are communication tools, and the level of detail and terminology can vary across laboratories and reporting systems. The process of interpreting the data itself is often time‐consuming, and variability in reporting practices can further exacerbate documentation workload. This combination of time burden and variability can delay report generation and reduce consistency for downstream communication and secondary use of EDX data.

Several groups have attempted to automate EDX reporting. Early rule‐based, probabilistic, and knowledge‐based systems were developed, but these tools lacked interactive input and relied on subjective grading of nerve damage [1]. Prior machine‐learning studies have trained classifiers to categorize NCS features consistent with specific conditions (e.g., carpal tunnel syndrome) [2]; however, the final clinical diagnosis remains a clinician's judgment integrating EDX findings with clinical context [3, 4]. Beyond classification, sequence‐to‐text approaches have been explored for drafting electrophysiology reports from structured test records [5]; however, these systems have typically been evaluated in constrained settings and may not readily extend across diverse protocols and reporting styles.

More recently, large language models (LLMs) have been explored for assisting clinical documentation, raising interest in their use for drafting portions of EDX reports. The INSPIRE line of work applied multi‐agent LLM pipelines to generate EDX reports and showed strong discrimination of normal versus abnormal studies in development. Yet in a randomized controlled trial the multi‐agent tool did not significantly enhance neurologists' performance and required substantial manual revision, limiting workflow integration [6]. Broader analyses of LLMs in medicine highlight both potential and pitfalls [7, 8], underscoring the need for designs that minimize hallucinations and reduce clinician editing burden. These observations have motivated hybrid designs that integrate LLMs with deterministic components and tightly constrain generation to produce standardized, clinician‐reviewable drafts with minimal manual revision.

To address these challenges, we pursued a hybrid rule–LLM approach that balances auditability and standardization. Deterministic rules can reliably translate measured EDX data into consistent structured descriptors, whereas drafting a clinician‐facing diagnostic interpretation often requires combining multiple findings into a coherent narrative with appropriate uncertainty; purely rule‐based narrative generation becomes difficult to scale and maintain, while unconstrained generation can introduce unsupported statements. The aim of this study was to develop and evaluate a hybrid rule–LLM pipeline for standardized EDX report drafting across two institutions.

2. Methods

2.1. Study Design and Setting

We conducted a retrospective study of EDX reports from two institutions: (i) a community primary‐care clinic where examinations were performed by the attending clinician, and (ii) a tertiary‐care rehabilitation medicine department where examinations were performed by residents and finalized under supervision of a board‐certified specialist.

The primary‐care cohort spanned April 1, 2025 to June 30, 2025; the tertiary‐care cohort spanned September 1, 2023 to February 29, 2024. Diagnostic devices differed across sites. Examination rooms maintained fixed temperature/humidity per local laboratory policy. Each laboratory used its routine age‐ and sex‐adjusted cutoff‐based reference values; no cross‐harmonization of reference tables was undertaken. Reporting followed the STROBE recommendations for observational studies.

2.2. System Overview and Workflow

The system was designed as a hybrid pipeline for EDX report drafting. Briefly, (i) report content was extracted from source files and normalized to a unified schema; (ii) a deterministic rule engine generated a structured Summary of abnormalities based on laboratory‐specific reference thresholds; and (iii) a label‐constrained LLM drafted the clinician‐facing diagnostic interpretation from a predefined diagnostic‐label inventory, with limited free text only when necessary.

The overall pipeline and user‐interface (UI) workflow of the system are shown in Figures [Link], [Link].

2.3. Participants and Case Unit

The analytic unit was the EDX report. Eligible examinations were those performed in patients aged ≥ 15 years. If a patient underwent multiple examinations within the study window, these were retained as independent cases.

2.4. Exclusion Criteria (Applied in Order)

Out‐of‐scope testing; examinations that included procedures beyond the prespecified scope (standard upper/lower‐limb NCS and limb/paraspinal needle EMG) were excluded because the deterministic Summary rules and diagnostic‐label inventory did not cover these protocols.

Central nervous system (CNS) lesion previously diagnosed or supported by imaging/clinical records. These examinations were excluded to avoid confounding by central lesions, as the study focus was peripheral neuromuscular EDX interpretation using standardized NCS/needle EMG patterns.

Entirely normal EDX report (no parameter exceeding the laboratory‐specific abnormal threshold in NCS or needle EMG). Because the pipeline deterministically outputs a normal summary and diagnostic interpretation when no abnormalities are detected, including normal reports would contribute limited information and would trivially inflate agreement metrics. We therefore restricted this proof‐of‐principle evaluation to abnormal, in‐scope reports. Expected pipeline behavior on normal reports is documented in the Supporting Information Methods (Table S4).

Abnormal but diagnostically noninformative examinations, determined by dual independent author review with consensus. Typical scenarios include artifacts or protocol insufficiency precluding interpretation, and minor/nonspecific abnormalities that do not support any diagnostic significance. We treated these as nondiagnostic studies and excluded them because they do not yield a reliable report‐level diagnostic interpretation against which agreement can be meaningfully assessed. Operational categories and illustrative examples used in this adjudication are summarized in the Supporting Information Methods (Table S5).

2.5. Rule‐Based Summary Module: Scope and Principle

In this module, the Summary records abnormalities observed in the EDX report verbatim and without clinical inference, strictly according to each laboratory's cutoff‐based reference values. The module applies prespecified rules to the raw measurements to generate report‐style natural language statements. Terminology and parameter usage were aligned with the American Association of Neuromuscular & Electrodiagnostic Medicine–International Federation of Clinical Neurophysiology glossary [9]. All rule‐based Summary were double‐checked against the raw measurement tables for all examinations.

2.6. NCS Abnormality Detection

For each pair of nerve/side/stimulation site/recording site, abnormality flags were assigned using laboratory‐specific thresholds: for motor NCS, compound motor action potential (CMAP) amplitude, distal latency, and conduction velocity; for sensory NCS, sensory nerve action potential (SNAP) amplitude, latency (onset or peak per laboratory convention), and conduction velocity; and for F‐wave and H‐reflex studies, latency. Responses were marked as absent when the relevant latency and/or amplitude fields were zero or left blank in the source report.

2.7. Segment Definition

For motor studies, within a given (nerve, side, stimulating site, recording site) chain, each adjacent pair of stimulation sites constituted one segment. For sensory studies, each recording route/site for a nerve constituted one segment (e.g., wrist vs. palm routes). Inching and other specialty studies were treated as multiple short segments along the specified path.

2.8. Conduction Block (CB)/Temporal Dispersion (TD) Proxy Descriptors

Because formal CB and TD determinations often require clinical context and area‐based waveform review, we reported simple proxy descriptors instead of asserting CB/TD. Intersegment amplitude drops of 30%–49% between adjacent stimulation sites (comparing the more proximal response with the immediately distal response) were recorded as “moderate drop,” and drops of ≥ 50% as “severe drop” [10, 11]. Intersegment duration increases of ≥ 30% were recorded as “duration increase” (tibial motor studies used a threshold of ≥ 100%) [12]. If the distal response was present but the immediately proximal response was absent, we conservatively treated the intersegment comparison as a ≥ 50% amplitude drop and reported a severe drop (study convention).

2.9. EMG Mapping and Abbreviations

Synonymous muscle names and abbreviations (e.g., ADQ ≡ ADM) were normalized to canonical labels. EMG attributes included insertional activity, fibrillation potentials and positive sharp waves, fasciculation potentials, motor unit action potential (MUAP) amplitude/duration/phases, and interference pattern.

2.10. EMG Abnormality Detection

Abnormalities were recorded as observed for each muscle: (i) MUAP analysis (amplitude, duration, phases); (ii) fibrillation potentials and positive sharp waves; (iii) fasciculation potentials; (iv) insertional activity; (v) recruitment; and (vi) interference pattern.

2.11. Evaluation of Summary Reporting

To quantify omissions in narrative reporting, we compared the human‐written Summary with the rule‐based Summary. For NCS, per examination we counted segments with $\ge 1$ abnormal parameter and computed recall (reporting sensitivity) for human versus rule‐based outputs; overcalls (normal segments described as abnormal) were enumerated separately. For EMG, per examination we counted muscles with abnormality and, within each such muscle, the number of abnormal items recorded; human versus rule‐based recall and overcall counts were summarized.

2.12. Artificial Intelligence (AI) Diagnostic Interpretation Generation

AI diagnostic interpretation was generated by a LLM (OpenAI o3‐2025‐04‐16) conditioned on four inputs: (i) the rule‐derived structured Summary; (ii) the normalized exam tables containing all parsed EDX measurements; (iii) the laboratory‐specific NCS reference values used to determine abnormality; and (iv) a minimal clinical note (age, sex, chief complaint [CC], and symptom duration). Although the UI accepts free‐form clinical notes, for the present evaluation we restricted inputs by design to age, sex, CC, and symptom duration; all other narrative content was disabled. The model was constrained to a prespecified inventory of diagnostic label formats, with free‐text fallback only when none of the listed formats applied. In such cases, the output was explicitly flagged as a free‐text fallback for clinician review. In the evaluated cohort, the free‐text fallback was not triggered (0 occurrences). The Supporting Information Methods provide the diagnostic‐label inventory and tuple schema (Table S1) and selected ontology mapping rules used for label normalization (Table S2), along with a description of the model interface and output constraints.

Although this final drafting step could be implemented with deterministic templates and rules, such an approach scales poorly as the diagnostic‐label inventory grows and as label combinations (e.g., multilabel cases, laterality merging, and optional site/level expressions) increase. We therefore used a constrained LLM as a language‐realization layer to (i) compose a coherent clinician‐facing diagnostic interpretation from structured evidence, (ii) enforce consistent formatting/terminology, and (iii) assemble controlled descriptors when needed (e.g., polyneuropathy phenotype descriptors composed from a limited token set; Supporting Information Methods S1), while prohibiting unsupported free‐text generation.

A separate UI mode that produces standardized, template‐based clinician comments was implemented and available but it was outside the scope of the present evaluation and was not analyzed.

2.13. Label Normalization and Ontology

Both the clinician's diagnostic interpretation and the AI diagnostic interpretation were normalized to a common label tuple: (family, structure, side, site), where side was coded as {right, left, bilateral, unspecified}. We kept bilateral statements as a single bilateral label to reflect how the diagnostic interpretation is written in practice. Parent–child relations within the ontology allowed partial structural matches during agreement scoring. Side handling used exact and reduced‐credit matches as specified in Section 2.14. “Lumbar radiculopathy” and “lumbosacral radiculopathy” were treated as the same structure for normalization.

2.14. Agreement Metric for Diagnostic Interpretation

Agreement between the AI‐generated and clinician diagnostic interpretation was quantified using the coverage–localization index (CLI) at the examination level and summarized across examinations (macro‐average). The CLI quantifies agreement in diagnostic labels after normalization to a shared ontology; it does not constitute an independent measure of diagnostic correctness against external ground truth. CLI combines coverage (whether the same diagnostic labels were selected) and Localization (whether the anatomic site/level matched when applicable) using a prespecified weighted sum (0.7 × coverage + 0.3 × localization). Sensitivity analyses using alternative weightings are provided in the Supporting Information Methods (Table S6). When an examination contained no labels eligible for localization, CLI was defined as coverage only.

Coverage was computed as an F1 score over sets of diagnostic labels based on (family, structure, side), ignoring site/level. To prevent double‐counting when the number of predicted and reference labels differed, labels were matched one‐to‐one to maximize overall agreement [13]. Agreement was scored as 1.0 for exact matches and 0.5 for a limited set of prespecified parent–child matches reflecting nerve‐branch naming granularity (e.g., “peroneal” vs. “deep peroneal”); otherwise, agreement was 0.

Localization assessed site/level agreement for label families where localization is clinically meaningful (e.g., mononeuropathy, radiculopathy, and plexopathy). Localization was scored as 1.0 for correct site/level, 0.5 when site/level was omitted, and 0.0 when site/level was incorrect. Full scoring details (including one‐to‐one matching and weighting) and the parent–child mapping table are provided in the Supporting Information Methods (Table S3) and Supporting Cases.

2.15. Safety‐Critical Errors

Independently of the index, we recorded a binary safety‐critical error flag (not included in Total): side inversion (left/right), normal/abnormal inversion.

2.16. Statistical Analysis

We summarized coverage, localization, and total as means with 95% confidence intervals (CIs) obtained by the bootstrap method [14]. Prespecified subgroup analyses included diagnostic family and clinical setting (primary vs. tertiary). Counts for exclusion categories and screening outcomes appear in Section 3.

2.17. Data Handling and Ethics

Source reports were de‐identified per institutional policy before analysis, and no patient identifiers were retained. This retrospective study was conducted under an Institutional Review Board (IRB) reliance agreement between Chungnam National University Hospital (CNUH; tertiary‐care) and a collaborating primary‐care clinic. The IRB of CNUH reviewed the protocol, and the requirement for informed consent was waived due to the retrospective design and full anonymization of data. All procedures were conducted in accordance with institutional guidelines and the Declaration of Helsinki.

3. Results

3.1. Study Cohorts

We screened 500 examinations across the two sites and included 332 abnormal, in‐scope EDX reports for analysis (Table 1). Out‐of‐scope protocols were uncommon overall (5/500 examinations, 1.0%). Inclusion was 241/295 in primary care and 91/205 in tertiary care, with exclusions due to CNS‐related diagnoses, normal reports, nondiagnostic studies, and out‐of‐scope protocols.

TABLE 1.

Screening and inclusion by site.

	Initial	Included	Excluded	CNS	Normal	Nondiagnostic	Out‐of‐scope
Primary	295	241	54	2	39	13	0
Tertiary	205	91	114	58	43	8	5
Overall	500	332	168	60	82	21	5

Note: Counts are number of examinations.

3.2. Diagnostic Interpretation Label Distribution

We summarized the distribution of diagnostic interpretation labels after standardizing label names. Across 332 included cases, 398 total labels were observed (k = 1 in 270 cases, k = 2 in 58 cases, k = 3 in 4 cases) (Table 2).

TABLE 2.

Diagnostic interpretation labels by frequency.

Label	Cases	Share of all labels (%)	Prevalence per case (%)
Median–Wrist	134	33.7	40.4
Lumbosacral radiculopathy	107	26.9	32.2
Peripheral polyneuropathy	42	10.6	12.7
Ulnar–Elbow	39	9.8	11.7
Cervical radiculopathy	23	5.8	6.9
Brachial plexus injury	6	1.5	1.8
Lateral femoral cutaneous neuropathy	4	1.0	1.2
Superficial peroneal	4	1.0	1.2
Common peroneal‐Knee	3	0.8	0.9
Median–Elbow	3	0.8	0.9
Motor neuron disease	3	0.8	0.9
Myopathy	3	0.8	0.9
Tibial–Ankle	3	0.8	0.9
Others (aggregated)	24	6.0	7.2

Note: Percentages are relative to all labels (398) or all cases (332).

3.3. Summary Reporting Recall

Summary reporting recall was generally high across all examinations, with a higher frequency of omissions observed in tertiary care settings (Table 3).

TABLE 3.

Summary reporting recall (macro recall, per examination).

	Primary (n = 241)	Tertiary (n = 91)	Overall (n = 332)
NCS recall (Human)	99.5	91.3	96.9
NCS recall (Rule)	100.0	100.0	100.0
EMG recall (Human)	97.6	98.8	97.9
EMG recall (Rule)	100.0	100.0	100.0

Note: Values are percentages.

3.4. Diagnostic Interpretation Agreement

Agreement between AI and clinicians was high overall, and there were no safety‐critical errors. Small decrements in CLI agreement were observed in the tertiary‐care cohort (Table 4).

TABLE 4.

Agreement between AI and clinician diagnostic interpretation.

	Coverage	Localization	Total
Primary (n = 241)	98.7 (97.4–99.8)	97.4 (95.7–98.8)	98.3 (96.9–99.4)
Tertiary (n = 91)	93.8 (89.0–97.6)	90.0 (85.2–94.3)	92.6 (88.1–96.5)
Overall (n = 332)	97.3 (95.7–98.7)	95.4 (93.5–97.0)	96.7 (95.1–98.1)

Note: Values are mean percentage (95% confidence interval).

3.5. Complexity–Performance Relationship

As complexity increased—operationalized by the total count of abnormal items per examination—both the human omission rate in Summary and the agreement of diagnostic interpretation worsened monotonically (Table 5).

TABLE 5.

Performance by examination complexity.

Complexity	n	Total abnormal count, mean (min–max)	Human omission rate (%)	Mean total agreement (%)
Low	141	2.74 (1–4)	5.7	98.8
Mid	100	6.47 (5–8)	6.0	97.4
High	91	16.23 (9–44)	31.9	92.8

Note: Complexity tertiles were defined by total abnormal count as low ≤ 4, mid 5–8, and high ≥ 9.

Concordantly, cases with a single diagnostic interpretation label were near‐perfect, whereas those with multiple labels showed a graded decline in coverage, localization, and total (Table 6).

TABLE 6.

Agreement by number of diagnostic interpretation labels.

k (labels)	n	Coverage	Localization	Total
1	271	99.3 (98.9–99.7)	97.2 (96.3–98.0)	98.6 (98.2–98.9)
2	57	89.2 (85.5–92.6)	88.3 (84.6–91.8)	88.9 (85.3–92.1)
3	4	83.3 (69.0–97.5)	70.8 (51.3–89.6)	79.6 (66.7–92.5)

Note: Values are percentages (95% CI).

The magnitude of the complexity effect was clinically meaningful: compared with Low, High complexity increased the human omission rate by +26.2 percentage points and reduced Total agreement by −6.1 percentage points, with bootstrap 95% CIs (Table 7).

TABLE 7.

Complexity contrast.

Outcome	High minus low (pp)	95% CI (pp)
Human omission rate (Summary)	26.2	16.2 to 36.8
Total agreement	−6.1	−10.7 to −2.0

Note: Differences are reported in percentage points (pp).

3.6. Overcalls

In the Summary, overcalls were rare: no overcalls occurred in the primary‐care cohort, whereas eight overcalls were observed across 2332 evaluated NCS segments in four of 91 tertiary‐care examinations. Because this section targets completeness, we report macro recall as the primary metric. Incorporating overcalls quantitatively would introduce design‐dependent choices with limited impact given the rarity, so we omit them from the primary metric. In diagnostic interpretation, extra labels are penalized through precision in the F1 with one‐to‐one matching, so overcalls reduce coverage by design.

4. Discussion

In this cohort restricted to abnormal, in‐scope EDX reports, the hybrid AI pipeline produced a structured Summary and a constrained diagnostic interpretation with high agreement to clinician diagnostic interpretation. Normal reports were not included in agreement analyses because the pipeline deterministically outputs a normal report when no abnormalities are detected and would trivially inflate agreement; nondiagnostic examinations were excluded. These results should therefore be interpreted as proof‐of‐principle for documentation drafting rather than evidence of diagnostic accuracy or clinical impact.

EDX interpretation depends on more than numeric thresholds. Technical quality (e.g., temperature optimization, electrode placement, adequate stimulation, and artifact recognition) and appropriateness of the test selection relative to the referral question materially influence how findings are weighed and communicated. Accordingly, apparent omissions in clinician narratives may reflect deliberate judgment in borderline or technically limited scenarios, rather than simple documentation errors. These considerations help explain why fully autonomous EDX interpretation remains challenging and motivate a constrained, clinician‐reviewable documentation support approach rather than unsupervised automation.

We therefore framed the rule‐based component as a descriptive layer that standardizes how raw values are transcribed against laboratory reference cutoffs, while avoiding clinical inference. The generative component was used only to draft a readable diagnostic interpretation from a predefined diagnostic‐label inventory, which reduces open‐ended generation and helps enforce consistent terminology. In principle, a deterministic label‐selection module could be constructed for this final step; we used a constrained LLM to flexibly compose clinician‐facing sentences and to handle heterogeneous combinations of labels and laterality patterns without expanding an ever‐growing set of hand‐written templates.

Compared with earlier expert systems and machine‐learning classifiers, our evaluation focuses on report drafting and structured label agreement rather than detection from raw signals. Prior work illustrates both the promise and limitations of automated reporting, including the need for clinician review and the risk that strong claims about workflow benefit are not realized without measured usability outcomes [1, 3, 5, 6, 8].

Several limitations are important. First, spectrum bias remains because diagnostically ambiguous or nondiagnostic reports were excluded, and agreement estimates reflect abnormal, in‐scope reports, which would inflate agreement in an imbalanced stream. Second, we did not measure workflow outcomes (time‐to‐completion, edit burden, or clinician acceptability), so any efficiency gains remain unproven. Third, external validity is limited because only two institutions were evaluated, and repeated examinations from the same patient may introduce within‐patient correlation that could narrow bootstrap CIs. Finally, the current scope does not address upstream quality control or whether the performed protocol adequately answers the referral question, both of which are central to clinical EDX practice.

Future work should include prospective, multicenter validation on consecutive routine EDX reports, intentionally including normal and diagnostically ambiguous studies, and should quantify workflow impact using usability metrics such as time‐to‐completion, edit burden, and clinician acceptability ratings. Studies should also incorporate systematic auditing for safety‐relevant errors. In addition, broader structured clinical context, including symptom distribution, disease history, examination findings, and referral‐specific information, should be evaluated within a controlled‐generation framework to improve disambiguation of complex cases without increasing hallucination risk.

Author Contributions

Kang Hee Cho: writing – review and editing, conceptualization, methodology, investigation, visualization. Ja Young Choi: data curation, supervision, validation, project administration, formal analysis. Yesung Jung: conceptualization, investigation, writing – review and editing, writing – original draft, methodology, software, data curation, visualization, project administration. Min Kyun Sohn: methodology, validation, supervision, data curation.

Ethics Statement

The authors have nothing to report.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Figure S1: Four‐stage pipeline.

Figure S2: Interpreter screen where parsed exam tables are normalized and auto‐generated into the UI; after parsing, autofill notes creates rule‐based notes and Generate Summary produces the structured Summary.

Figure S3: Report panel showing rule‐based Summary and the pending action to generate the diagnostic interpretation.

Table S1: Diagnostic label tuple schema and controlled vocabulary used for scoring.

Table S2: Selected ontology mapping rules and examples used for label normalization.

Table S3: Parent–child mapping used for partial agreement in coverage.

Table S4: Pipeline outputs on a subset of normal EDX reports.

Table S5: Nondiagnostic categories and illustrative scenarios.

Table S6: Sensitivity analysis of CLI weighting.

Table S7: Representative discordant examples illustrating limitations from restricted clinical inputs.

Data Availability Statement

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.