Why this matters for Axiomera & Axiomera

The empirical proof that semantic classification and harmonization hold up at scale

This study is the strongest evidence to date behind the Axiomera platform's core premise: that binding messy EHR data to a formal semantic-temporal structure — entities, relationships, temporal components, and context — turns siloed warehouse records into analyzable clinical intelligence. The framework's semantic abstraction, hybrid concept extraction against medical ontologies, and FHIR R4 interfaces are the same building blocks Axiomera uses to classify, map, and enrich clinical data before harmonizing it across sources.

The scale is the point. Holding accuracy and fairness across 492,542 encounters at three structurally different sites — an academic medical center, a community hospital, and a multispecialty clinic — is exactly the multi-source, multi-context problem Axiomera's Domain Intelligence Layer is built to solve. The reported 75–95% functional retention when core components are adapted to other EHR platforms speaks directly to the vendor-agnostic design that lets the platform operate without moving data out of its domain.

For the clinical-facing sibling, the validated relationship discovery and temporal care-pathway modeling map onto population-health and longitudinal-outcomes use cases. If you want to see how this translates from paper to production, request a technical briefing.

Abstract

Electronic Health Record (EHR) data warehouses present significant challenges for clinical analytics due to heterogeneous data structures and unstructured clinical narratives. This paper introduces a unified Semantic Intelligence Framework that demonstrates how semantic-temporal representations of EHR data can systematically expose clinically validated relationships and care pathway archetypes inaccessible to conventional analytics approaches. Our framework integrates graph-based concept extraction using hybrid NLP approaches, statistically validated relationship discovery with rigorous multiple testing correction, and comprehensive temporal pattern analysis using bidirectional LSTM architectures with clinical attention mechanisms. Evaluated across three healthcare systems encompassing 492,542 patient encounters from 127,834 patients, the framework demonstrates associations with improved diagnostic coding accuracy (8.3 percentage points, p < 0.001), enhanced comorbidity detection (12.3 percentage points, p < 0.001), and increased clinical decision support relevance (15.4 percentage points, p < 0.001). Statistical validation identified 156 clinically significant relationships from 34.8 million candidate pairs, achieving a 63.2% clinical validation rate. The system maintains real-time performance with clinical decision support recommendations generated in 2.3 ± 0.7 seconds per encounter. Comprehensive algorithmic fairness assessment confirmed minimal bias across demographic groups, with all demographic parity differences below the 0.10 concern threshold. Our results indicate that semantic intelligence approaches can significantly enhance clinical analytics while maintaining robustness and fairness across diverse patient populations. The framework's modular architecture supports generalizability beyond the current Epic Clarity implementation to other EHR platforms through standardized semantic abstractions.

Keywords: semantic intelligence, healthcare data analytics, clinical decision support, natural language processing, temporal pattern analysis, algorithmic fairness, generalizable framework.

In this paper
  1. Introduction
  2. Related Work
  3. Methodology
  4. Experimental Design
  5. Results
  6. Discussion
  7. Conclusion
  8. Appendices
  9. References

1. Introduction

Healthcare organizations face persistent challenges in deriving actionable insights from heterogeneous Electronic Health Record (EHR) data. While EHR data warehouses provide comprehensive clinical information, they present substantial integration challenges across structured, unstructured, temporal, and workflow data domains [1]. Traditional query-based approaches often fail to capture the complex, context-rich nature of healthcare information, limiting their utility for comprehensive clinical intelligence [2].

Current artificial intelligence applications in healthcare remain fragmented, focusing on specific clinical tasks rather than unified clinical intelligence frameworks [3]. While natural language processing (NLP) advancements show promise [4], they typically operate in isolation without integration into broader clinical workflows. Similarly, temporal analysis methods [5, 6] often concentrate on specific disease conditions rather than providing holistic longitudinal intelligence.

This paper presents a unified Semantic Intelligence Framework that addresses these limitations through a generalizable approach to clinical data analytics. Our primary contribution is demonstrating that a unified semantic-temporal representation of EHR data can systematically expose clinically validated relationships and care pathway archetypes that are inaccessible to conventional SQL-based or siloed analytics approaches. The framework's core innovations include:

  1. A vendor-agnostic semantic abstraction \(S = (E, R, T, C)\) representing clinical entities, relationships, temporal components, and contextual meaning.
  2. Hierarchical concept extraction using hybrid NLP approaches combining ClinicalBERT with rule-based medical terminology recognition.
  3. Statistically validated relationship discovery with rigorous multiple testing correction and structured clinical expert review.
  4. Comprehensive temporal pattern analysis using bidirectional LSTM architectures with clinical attention mechanisms.

Our contributions include:

2. Related Work

2.1 Clinical Natural Language Processing

Recent advances in clinical NLP have enabled more sophisticated extraction of information from unstructured clinical notes [1]. Large language models specifically designed for EHR data have shown promising results in clinical concept extraction and relationship identification [7]. However, these approaches often operate in isolation without integration into comprehensive clinical intelligence frameworks that span the entire analytics pipeline from raw data to clinical insights.

2.2 Clinical Relationship Discovery

Statistical methods for discovering clinical relationships from EHR data have evolved significantly [8]. While methods like association rule mining and co-occurrence analysis have been applied, they often lack rigorous statistical validation and structured clinical expert review [9]. Our framework addresses these limitations through comprehensive multiple testing correction, effect size estimation, and systematic clinical validation processes that distinguish meaningful patterns from statistical artefacts.

2.3 Temporal Pattern Analysis

Temporal analysis in healthcare has focused primarily on specific clinical conditions or prediction tasks [5]. Recent approaches using deep learning architectures like LSTMs and transformers have shown promise but often lack clinical interpretability [6]. Our bidirectional LSTM architecture with clinical attention mechanisms provides both predictive accuracy and interpretable temporal insights, enabling understanding of which clinical events drive specific predictions.

3. Methodology

3.1 Semantic Framework Foundation

The Semantic Intelligence Framework is built upon a formal representation of clinical data as a measurable algebraic structure \(S = (E, R, T, C)\), where:

This semantic abstraction enables consistent representation of clinical information across different EHR systems and supports the framework's generalizability beyond specific implementations.

A unified semantic-temporal representation of EHR data can systematically expose clinically validated relationships and care pathway archetypes that are inaccessible to conventional SQL-based or siloed analytics approaches.

3.2 System Architecture

The framework employs a five-layer distributed processing architecture specifically designed for clinical data warehouses (Figure 1). Each layer addresses specific aspects of clinical intelligence while maintaining interoperability through standardized interfaces.

Five-layer distributed architecture of the semantic intelligence framework, from data-warehouse integration through concept extraction, relationship discovery, and temporal analysis up to clinical decision support.
Figure 1. Unified Semantic Intelligence Framework: generalizable architecture for clinical data analytics.

3.3 Data Integration and Quality Assurance

The framework integrates with clinical data warehouses through standardized SQL interfaces for batch analytics and FHIR R4 APIs for real-time clinical decision support. Data quality metrics include completeness \(C(D)\) and consistency measures:

$$ C(D) = \frac{N_{\text{complete}}}{N_{\text{total}}} \times 100\% $$
$$ \text{Consistency}(T_1, T_2) = \frac{|R_1 \cap R_2|}{|R_1 \cup R_2|} $$

Table 1 summarizes data integration quality metrics across core clinical tables.

Table 1. Clinical data integration summary with quality metrics.
Data sourceRecord countCompleteness (%)Consistency scoreProcessing rate (rec/s)
Patient Encounters3,847,29399.70.9822,847
Clinical Notes834,26794.20.9671,234
Diagnostic Codes847,32998.10.9913,421
Medication Orders1,923,84796.80.9752,156
Laboratory Results12,347,89297.30.9894,672

3.4 Clinical Concept Extraction

The concept extraction module combines ClinicalBERT embeddings with rule-based medical terminology recognition and ontology mapping. Entity confidence is computed using a weighted probabilistic model:

$$ P(\text{entity}\mid\text{context}) = w_1 \times P_{\text{BERT}} + w_2 \times P_{\text{ont}} + w_3 \times P_{\text{clin}} $$

where weights are optimized as \(w_1 = 0.45\), \(w_2 = 0.35\), \(w_3 = 0.20\) through cross-validation. This hybrid approach balances the contextual understanding of transformer models with the precision of rule-based systems and clinical knowledge bases.

3.5 Statistical Relationship Discovery

Relationship discovery employs rigorous statistical testing with multiple comparison correction (Figure 2). The candidate generation process begins with:

$$ N_{\text{candidates}} = \binom{n}{2} = \frac{n(n-1)}{2} $$
Relationship discovery pipeline flowing from millions of candidate concept pairs through Fisher's exact testing, effect-size estimation, and multiple-testing correction to clinically validated associations.
Figure 2. Statistically rigorous relationship discovery pipeline: from candidate pairs to clinically validated associations.

Statistical significance is assessed using Fisher's exact test, with effect sizes measured via Cramér's V:

$$ V = \sqrt{\frac{\chi^2}{n \times \min(\text{rows}-1,\ \text{cols}-1)}} $$

Multiple testing correction employs the Benjamini-Hochberg procedure [10] to control false discovery rate at \(q = 0.05\). This rigorous approach ensures that identified relationships represent meaningful clinical associations rather than chance occurrences.

3.6 Temporal Pattern Analysis

The temporal analysis framework processes patient trajectories \(T = \{(t_1, e_1), (t_2, e_2), \ldots, (t_k, e_k)\}\) using dynamic time warping with clinical distance metrics:

$$ \text{DTW}(T_1, T_2) = \min \sum_{(i,j)\in P} d(e_{1i}, e_{2j}) + \lambda \times |t_i - t_j| $$

Bidirectional LSTM architectures with clinical attention mechanisms enable comprehensive temporal pattern recognition while maintaining interpretability through attention weights that highlight clinically relevant events.

3.7 Clinical Validation Framework

A multidisciplinary expert panel comprising board-certified physicians, clinical informaticists, and medical coding specialists conducted structured validation using standardized assessment protocols. Inter-rater reliability was measured using Cohen's kappa (\(\kappa\)):

$$ \kappa = \frac{p_o - p_e}{1 - p_e} $$

where \(p_o\) represents observed agreement and \(p_e\) represents expected agreement. This rigorous validation process ensures clinical relevance and practical utility of discovered patterns.

4. Experimental Design

4.1 Study Design and Healthcare Systems

We conducted a multi-site retrospective cohort study with prospective validation across three healthcare systems:

The study encompassed 127,834 patients with 492,542 clinical encounters over 12 months. IRB approval was obtained with waiver of informed consent for de-identified data analysis.

4.2 Implementation Protocol

Implementation followed a phased approach emphasizing modularity and generalizability:

  1. Phase I: Baseline establishment and clinical data warehouse integration via standardized interfaces.
  2. Phase II: Containerized microservices deployment using Kubernetes orchestration.
  3. Phase III: Clinical concept extraction model training with site-specific customization.
  4. Phase IV: Statistical relationship discovery and temporal pattern analysis.
  5. Phase V: Clinical decision support integration and validation.

4.3 Clinical Expert Validation

A multidisciplinary expert panel conducted blinded validation using standardized assessment protocols. Inter-rater reliability achieved Cohen's kappa (\(\kappa\)) values exceeding 0.85 across all specialties after standardized training (Table 2).

Table 2. Clinical expert validation panel composition and inter-rater reliability.
SpecialtyExperience (years)Validation casesCohen's κ
Internal Medicine8.3 ± 2.12,8470.89
Emergency Medicine6.7 ± 1.81,9230.87
Clinical Informatics12.4 ± 3.23,4210.92
Clinical Pharmacy9.1 ± 2.72,1560.85
Medical Coding11.8 ± 4.11,6740.91

4.4 Statistical Analysis

The comprehensive statistical analysis plan included power analysis, multiple comparison correction, and clinical significance assessment. Sample size calculations ensured adequate power (> 99%) for all primary analyses. The Benjamini-Hochberg procedure controlled false discovery rate at \(q = 0.05\). All reported improvements represent observed associations from retrospective analysis, with causal impact requiring prospective validation.

5. Results

5.1 System Performance

The framework demonstrated significant performance improvements across all processing components (Figure 3). Entity recognition processing time improved from 12.4 ± 1.8 seconds to 2.3 ± 0.4 seconds (81.5% improvement, p < 0.001). Overall system performance improved from 153.6 ± 12.7 seconds to 28.9 ± 2.8 seconds (81.2% improvement, p < 0.001).

System performance charts comparing baseline and optimized processing times across entity recognition and overall pipeline stages, with resource utilization metrics.
Figure 3. Comprehensive system performance analysis: processing efficiency and resource optimization.

5.2 Clinical Concept Extraction

The concept extraction system processed 834,267 clinical notes, identifying 8,347 unique clinical concepts with overall precision of 0.753 ± 0.028 and recall of 0.726 ± 0.032 (Figure 4). Ontology mapping achieved success rates of 91.3% for RxNorm medications and 85.7% for ICD-10-CM diagnostic concepts.

Concept extraction performance panels showing precision and recall by clinical domain alongside RxNorm and ICD-10-CM ontology mapping success rates.
Figure 4. Clinical concept extraction performance: multi-domain analysis and validation.

5.3 Relationship Discovery

From 34.8 million candidate concept pairs, the framework identified 156 clinically validated relationships (63.2% validation rate) through rigorous statistical filtering and expert review (Figure 5). Relationships were distributed across cardiovascular (23.7%), endocrine (18.6%), infectious disease (15.4%), and other clinical domains.

5.3.1 Clinical Vignettes

Example discovered relationship: The framework identified a statistically significant association between hemoglobin A1c variability and subsequent cardiovascular medication adjustments (p < 0.001, Cramér's V = 0.24), validated by endocrinology experts as clinically meaningful for diabetes management.

Additional discovery: Analysis revealed a significant relationship between depression severity documentation patterns and antidepressant dosage adjustment frequency (p < 0.001, Cramér's V = 0.19), highlighting opportunities for improved mental health monitoring.

Relationship discovery results showing the funnel from 34.8 million candidate pairs to 156 validated associations and their distribution across clinical domains.
Figure 5. Clinically validated relationship discovery: statistical rigor and expert consensus.

5.4 Temporal Pattern Analysis

The temporal analysis identified 89 statistically significant patterns from 492,542 clinical encounters (Figure 6). Common patterns included diabetes management sequences (23.4% of relevant trajectories) and hypertension monitoring patterns (19.7%). The bidirectional LSTM architecture achieved 0.784 ± 0.041 accuracy for 30-day clinical event prediction.

5.4.1 Exemplar Care Pathway

Analysis revealed a common diabetes management pathway where patients typically progress from initial diagnosis and lifestyle recommendations to medication initiation within 90 days, followed by dosage adjustments based on laboratory monitoring. This pattern was validated by clinical experts as representing standard best-practice care sequences.

Temporal pattern analysis visualizing longitudinal patient trajectories, common care-pathway archetypes, and 30-day event prediction accuracy from the bidirectional LSTM.
Figure 6. Temporal pattern analysis and care pathway intelligence: longitudinal patient trajectory modeling.

5.5 Clinical Decision Support Effectiveness

Integration with clinical decision support systems demonstrated significant associations with improved clinical outcomes (Figure 7). Diagnostic coding accuracy showed an association with improvement of 8.3 ± 1.5 percentage points across all sites (p < 0.001), comorbidity detection demonstrated an association with improvement of 12.3 ± 1.6 percentage points (p < 0.001), and decision support relevance showed an association with improvement of 15.4 ± 2.1 percentage points (p < 0.001).

Note: These associations represent observed improvements from retrospective analysis. Causal impact requires prospective controlled evaluation.

Clinical decision support effectiveness bar charts showing percentage-point improvements in diagnostic coding accuracy, comorbidity detection, and decision-support relevance across all sites.
Figure 7. Clinical decision support effectiveness: multi-dimensional performance assessment (all associations represent non-causal relationships).

5.6 Scalability and Quality Assurance

The system demonstrated linear scalability up to 50 concurrent users with acceptable performance degradation (Figure 8). Quality metrics remained stable throughout the 12-month evaluation period, with F1-scores within statistical control limits across all clinical domains.

Enterprise performance panels showing linear scalability to 50 concurrent users, stable F1-scores within statistical control limits, and governance metrics over 12 months.
Figure 8. Enterprise-grade system performance: scalability, quality assurance, and governance.

5.7 Algorithmic Fairness

Comprehensive bias assessment demonstrated acceptable fairness across demographic groups (Figure 9). Demographic parity differences remained below the 0.10 threshold for all racial/ethnic comparisons, and equalized opportunity metrics showed minimal variation across age and gender groups.

5.7.1 Fairness Governance

The framework incorporates continuous fairness monitoring with predefined alert thresholds. Automated reporting to clinical governance committees ensures timely intervention when bias metrics indicate potential disparities in care delivery. Monthly ethics committee reviews provide oversight for all algorithmic fairness measures.

Algorithmic fairness assessment showing demographic parity and equalized opportunity metrics across racial, age, and gender groups, all below the 0.10 concern threshold.
Figure 9. Comprehensive algorithmic fairness assessment: multi-dimensional bias analysis and governance.

6. Discussion

Our Semantic Intelligence Framework addresses critical limitations in current healthcare analytics approaches by integrating graph-based concept extraction, statistically validated relationship discovery, and comprehensive temporal pattern analysis within a unified, generalizable architecture. The demonstrated associations with improved diagnostic accuracy, comorbidity detection, and clinical decision support relevance represent substantial advancements over traditional methods.

The framework's ability to identify 156 clinically validated relationships from 34.8 million candidate pairs highlights the power of combining statistical rigor with clinical expertise. This approach mitigates the risk of identifying spurious correlations while capturing meaningful clinical associations that might otherwise remain undiscovered. The inclusion of clinical vignettes and exemplar care pathways enhances interpretability and practical utility.

The temporal analysis capabilities provide unprecedented insights into patient trajectories and care pathways. The identification of 89 statistically significant temporal patterns enables more proactive clinical interventions and personalized care planning. The bidirectional LSTM architecture with clinical attention mechanisms represents a significant advancement in healthcare temporal analytics by maintaining both predictive accuracy and clinical interpretability.

6.1 Limitations and Future Work

Several limitations warrant consideration in interpreting these results. The 12-month evaluation period, while sufficient for initial validation, may not capture long-term system performance and user adaptation. The three-site implementation, while providing valuable multi-site validation, has limited geographic and organizational diversity. All reported improvements represent associations from retrospective analysis, with causal impact requiring prospective controlled evaluation.

Future work should focus on:

6.2 Generalizability Beyond Current Implementation

While our current implementation specifically targets Epic Clarity data warehouses, the framework's modular architecture and semantic abstraction \(S = (E, R, T, C)\) support generalizability across EHR platforms (Figure 10). Key design elements enabling cross-platform adaptation include:

Preliminary analysis indicates that core framework components maintain 75–95% of their functionality when adapted to other major EHR platforms, with primary adaptation effort focused on terminology mapping and workflow integration.

Cross-platform generalizability analysis showing 75 to 95 percent functional retention when core framework components are adapted to other major EHR platforms.
Figure 10. Framework generalizability analysis: cross-platform adaptability and performance consistency.

7. Conclusion

The Semantic Intelligence Framework represents a significant advancement in healthcare data analytics through its unified approach to clinical concept extraction, statistically validated relationship discovery, and comprehensive temporal pattern analysis. By demonstrating that semantic-temporal representations of EHR data can systematically expose clinically validated relationships and care pathway archetypes inaccessible to conventional analytics, the framework provides a generalizable foundation for next-generation clinical intelligence systems.

The observed associations with improved diagnostic accuracy (8.3 percentage points), comorbidity detection (12.3 percentage points), and clinical decision support relevance (15.4 percentage points) highlight the framework's potential to enhance clinical workflows and patient care. The system's scalability, performance optimization, comprehensive fairness assessment, and governance processes ensure practical utility across diverse healthcare settings.

As healthcare organizations continue to grapple with the challenges of deriving actionable insights from complex EHR data, semantic intelligence approaches offer a promising path forward. The integration of statistical rigor, clinical expertise, advanced computational methods, and a generalizable architecture enables more intelligent, contextual, and effective healthcare analytics that can adapt to evolving clinical needs and technological landscapes.

Future research should focus on longitudinal validation across diverse clinical settings, prospective evaluation of causal impact, and continued expansion of the framework's capabilities through integration with emerging AI technologies and clinical data standards.

Acknowledgments

The authors thank the clinical expert validation panel and participating healthcare systems for their contributions to this research. Special thanks to the technical teams involved in system implementation and validation. This research was supported by VMC MAR COM Inc. DBA HeyDonto.

8. Appendices

A. Statistical Analysis Plan

The comprehensive statistical analysis plan included power analysis, multiple comparison correction, and clinical significance assessment. Sample size calculations ensured adequate power (> 99%) for all primary analyses. The Benjamini-Hochberg procedure controlled false discovery rate at \(q = 0.05\). All statistical tests were two-sided with \(\alpha = 0.05\) determining statistical significance.

B. Implementation Details

Detailed implementation specifications including container orchestration configurations, specific SQL extraction patterns, hyperparameter settings, and API integration details are available in the online supplementary materials. The modular architecture documentation provides guidance for adaptation to different EHR platforms and clinical environments.

C. Ethical Considerations and Fairness Governance

The framework incorporates comprehensive privacy protections, including HIPAA-compliant de-identification, role-based access controls, and algorithmic fairness monitoring. All research activities received IRB approval with appropriate safeguards for patient data protection. The fairness governance framework includes continuous monitoring, automated alerting for bias detection, and regular ethics committee review to ensure equitable system performance across all patient populations.

Frequently asked questions

What is a semantic intelligence framework for clinical data warehouses?

It is a unified system that represents EHR data as a semantic-temporal structure \(S = (E, R, T, C)\) — entities, relationships, temporal components, and context — so that graph-based concept extraction, statistically validated relationship discovery, and temporal pattern analysis all operate on the same abstraction. In this study it was evaluated across 492,542 patient encounters from 127,834 patients at three healthcare systems.

How large was the validation study?

The framework was evaluated across three healthcare systems encompassing 492,542 clinical encounters from 127,834 patients over 12 months. It processed 834,267 clinical notes and screened 34.8 million candidate concept pairs, from which 156 clinically validated relationships were confirmed at a 63.2% validation rate.

Does the framework move patient data out of the source system?

No. It integrates with clinical data warehouses through standardized SQL interfaces for batch analytics and FHIR R4 APIs for real-time decision support, operating on data in place. The study reports HIPAA-compliant de-identification, role-based access controls, and IRB approval with waiver of informed consent for de-identified analysis.

Do the reported improvements prove the framework causes better outcomes?

No. The authors are explicit that all reported improvements — including +8.3pp diagnostic coding accuracy, +12.3pp comorbidity detection, and +15.4pp decision-support relevance — represent observed associations from retrospective analysis. Establishing causal impact requires prospective controlled evaluation.

References

  1. X. Yang, A. Chen, N. PourNejatian, H. C. Shin, K. E. Smith, C. Parisien, … and Y. Wu, "A large language model for electronic health records," Nature Partner Journal Digital Medicine, vol. 5, no. 1, p. 194, 2022.
  2. Z. Yu, C. Peng, X. Yang, C. Dang, P. Adekkanattu, B. G. Patra, … and Y. Wu, "Identifying social determinants of health from clinical narratives: A study of performance, documentation ratio, and potential bias," Journal of Biomedical Informatics, vol. 153, p. 104642, 2024.
  3. S. A. Alowais, S. S. Alghamdi, N. Alsuhebany, T. Alqahtani, A. I. Alshaya, S. N. Almohareb, … and H. A. Badreldin, "Revolutionizing healthcare: The role of artificial intelligence in clinical practice," BMC Medical Education, vol. 23, no. 1, p. 689, 2023.
  4. S. Han, R. F. Zhang, L. Shi, R. Richie, H. Liu, A. Tseng, … and F. R. Tsui, "Classifying social determinants of health from unstructured electronic health records using deep learning-based natural language processing," Journal of Biomedical Informatics, vol. 127, p. 103984, 2022.
  5. L. Wang, J. Lakin, C. Riley, Z. Korach, L. N. Frain, and L. Zhou, "Disease trajectories and end-of-life care for dementias: Latent topic modeling and trend analysis using clinical notes," in AMIA Annual Symposium Proceedings, 2018, pp. 1056–1065.
  6. D. Zhang, J. Thadajarassiri, S. Cansu, and E. Rundensteiner, "Time-aware transformer-based network for clinical notes series prediction," in Machine Learning for Healthcare Conference, 2020, pp. 566–588.
  7. S. Yang, P. Varghese, E. Stephenson, K. Tu, and J. Gronsbell, "Machine learning approaches for electronic health records phenotyping: A methodical review," Journal of the American Medical Informatics Association, vol. 30, no. 2, pp. 367–381, 2023.
  8. B. Romanowski, A. Ben Abacha, and Y. Fan, "Extracting social determinants of health from clinical note text with classification and sequence-to-sequence approaches," Journal of the American Medical Informatics Association, vol. 30, no. 8, pp. 1448–1455, 2023.
  9. S. S. Band, A. Yarahmadi, C. C. Hsu, M. Biyari, M. Sookhak, R. Ameri, … and H. W. Liang, "Application of explainable artificial intelligence in medical health: A systematic review of interpretability methods," Informatics in Medicine Unlocked, vol. 40, p. 101286, 2023.
  10. Y. Benjamini and Y. Hochberg, "Controlling the false discovery rate: A practical and powerful approach to multiple testing," Journal of the Royal Statistical Society: Series B (Methodological), vol. 57, no. 1, pp. 289–300, 1995.
Cite this paper

Nehzati, M. (2026). Semantic Intelligence Framework for Clinical Data Warehouses: A Generalizable Approach to Graph-Based Concept Extraction, Statistically Validated Relationship Discovery, and Temporal Pattern Analysis. Axiomera Research (paper under peer review, under peer review). https://axiomera.com/blog/semantic-intelligence-clinical-data-warehouses