Abstract
The extraction of phenotype information which is naturally contained in electronic health records (EHRs) has been found to be useful in various clinical informatics applications such as disease diagnosis. However, due to imprecise descriptions, lack of gold standards and the demand for efficiency, annotating phenotypic abnormalities on millions of EHR narratives is still challenging. In this work, we propose a novel unsupervised deep learning framework to annotate the phenotypic abnormalities from EHRs via semantic latent representations. The proposed framework takes the advantage of Human Phenotype Ontology (HPO), which is a knowledge base of phenotypic abnormalities, to standardize the annotation results. Experiments have been conducted on 52,722 EHRs from MIMIC-III dataset. Quantitative and qualitative analysis have shown the proposed framework achieves state-of-the-art annotation performance and computational efficiency compared with other methods.
Introduction
Electronic Health Records (EHRs) are a critical resource for clinical research, yet much of the phenotypic information—observable physical or biochemical characteristics—remains trapped in unstructured text. Extracting these phenotypes is difficult because clinical notes often use imprecise language, abbreviations, or implicit descriptions. Furthermore, the lack of large-scale "gold standard" labeled datasets makes supervised learning prohibitively expensive. Traditional tools like MetaMap or keyword-based searches often struggle with semantic nuances or suffer from extreme computational inefficiency when applied to large datasets.
Methodology
The proposed framework utilizes an unsupervised Auto-encoder architecture to learn semantic latent representations of EHR text. By mapping clinical narratives into a shared space with the Human Phenotype Ontology (HPO), the model can identify abnormalities without requiring manual labels.
The system employs a Transformer-based encoder ($E$) to generate latent vectors and a generator ($G$) to reconstruct the text, ensuring semantic fidelity. A key innovation is the inclusion of a CNN-based classifier ($D$) that enforces "distinct priors," forcing the model to distinguish between different phenotypic categories in the latent space. The model is "taught" the meaning of phenotypes by training on HPO term definitions and hierarchical structures.
Experimental Setup
The framework was evaluated using 52,722 discharge summaries from the MIMIC-III dataset. To benchmark performance, the authors developed a "silver standard" by mapping ICD-9 diagnosis codes to HPO terms using the OMIM database. The study focused on 24 general phenotypic categories and utilized nearly 13,800 HPO subclasses for training.
Results and Discussion
Performance and Efficiency
The proposed model achieved an F1 score of 0.6383, significantly outperforming traditional annotators like MetaMap (0.5576) and NCBO Annotator (0.5659). Beyond accuracy, the framework demonstrated remarkable computational efficiency.
As shown in the comparison of processing times, the proposed method annotated the entire 52,722-record dataset in just 40.2 minutes, whereas MetaMap required approximately 22 days. This speedup is attributed to the parallel processing capabilities of the Transformer and CNN architectures.
Semantic Understanding
Qualitative analysis highlighted the model's ability to recognize "implicit" phenotypes. For example, it correctly identified "Abnormality of blood-forming tissues" from a CT scan description, a task where syntax-dependent baseline methods failed. This confirms the model's ability to prioritize semantics over literal keyword matching.
Conclusion
This research presents a scalable, unsupervised solution for the automated annotation of phenotypic abnormalities in EHRs. By leveraging ontologies as a source of knowledge and utilizing semantic latent representations, the framework achieves state-of-the-art performance while being orders of magnitude faster than existing tools. Future work will aim to extend the model to the full HPO vocabulary and explore its utility in other domains.
Summary updated