Artificial intelligence processing electronic health records to identify commonalities and comorbidities cluster at Immuno Center Humanitas

Pierandrea Morandini, Maria Elena Laino, Giovanni Paoletti, Alessandro Carlucci, Tobia Tommasini, Giovanni Angelotti, J. Pepys, Giorgio Walter Canonica, Enrico Heffler, Victor Savevski, Francesca PuggioniClinical 2022 · DOI

TL;DR

This paper introduces an AI-powered approach using Natural Language Processing (NLP) to analyze electronic health records (EHRs) and identify commonalities and comorbidities among patients at an Immuno Center. By extracting structured information from unstructured clinical notes, the study reveals previously unrecognized patient clusters and offers insights into more efficient, multidisciplinary treatment strategies.

What Problem Does This Paper Address?

Chronic inflammatory conditions, like asthma, often come with multiple comorbidities, necessitating a complex, multidisciplinary approach to treatment. Understanding the intricate links between a primary disease and its associated conditions is crucial for effective patient management. However, a significant challenge lies in extracting this valuable information from the vast amount of unstructured data within electronic health records (EHRs), which are typically written in narrative, free-text format by clinicians. This paper aims to bridge this gap by demonstrating how NLP can be leveraged to systematically identify pathologies, patient phenotypes, and treatment patterns from these free-text notes, thereby enriching real-world evidence and facilitating a more holistic understanding of patient health.

Methods

The core of this study revolves around an NLP-based process for knowledge discovery from EHRs. The researchers employed Regular Expressions (RegEx) for marker extraction, a technique well-suited for identifying specific patterns and entities within text. This approach was chosen for its efficiency in handling the structured medical terminology and abbreviations found in clinical notes.

The process involved several key steps:

  • Data Collection: Clinical notes from the Allergy Department of Humanitas Research Hospital, spanning from January 2017 to September 2020, were retrospectively collected. Only records from patients who provided consent were included.
  • Data Selection and Pre-processing: Clinical notes were processed to handle various layout structures, with information specifically extracted from anamnesis (for comorbidities), conclusions (for final diagnoses), and therapy paragraphs. A normalization step was performed to clean the text data, converting characters to ASCII and removing HTML special characters.
  • Marker Extraction: RegEx patterns were developed to identify specific pathologies and therapies (referred to as "entities") within the cleaned text.
  • Clustering: A k-modes algorithm was used to group patients based on their identified comorbidities and diagnoses. This algorithm is a variation of k-means, specifically designed for categorical data, making it suitable for the binary flags representing the presence or absence of conditions. The optimal number of clusters was determined using the elbow method and silhouette analysis.

Datasets

The study utilized a dataset of clinical notes from the Allergy Department of Humanitas Research Hospital, collected retrospectively from January 2017 to September 2020. This dataset comprised 3,162 patients and 7,827 visits, with 2,057 patients ultimately enrolled after screening. The paper does not introduce a new benchmark dataset but rather uses real-world clinical data.

Key Insights and Ideas

  • NLP as a Powerful Tool for EHR Analysis: The study demonstrates that NLP, particularly using RegEx, can effectively extract structured information from unstructured clinical narratives, unlocking valuable insights that might otherwise remain hidden.
  • Identification of Patient Clusters: The k-modes clustering algorithm successfully grouped patients based on their comorbidity profiles, revealing distinct patient clusters with shared conditions. This moves beyond simply listing comorbidities to understanding how they co-occur.
  • Comorbidities as a Unifying Feature: The research highlights that comorbidities are not isolated occurrences but often form interconnected patterns, suggesting underlying shared pathophysiological mechanisms.
  • Potential for Multidisciplinary Care Optimization: By identifying these clusters, the study suggests that a more integrated, multidisciplinary approach to treatment might be beneficial, potentially leading to more efficient resource utilization and improved patient outcomes.
  • Reduced Corticosteroid Prescription: A notable finding was that the volume of systemic corticosteroid prescriptions at the Immuno Center was generally lower compared to patients' previous treatment regimens, with exceptions for COPD and angioedema.

Key Results

  • Frequent Comorbidities: The three most frequent comorbidities identified were asthma, rhinitis, and urticaria.
  • Patient Cohort with Comorbidities: Out of 2,057 patients, 991 suffered from at least one of the considered comorbidities.
  • Key Comorbidity Clusters: Significant co-occurrences were observed, including:
  • Oral allergy syndrome and urticaria (131 patients)
    • Angioedema and urticaria (105 patients)
    • Rhinitis and asthma (227 patients)
  • Marker Extraction Performance: The RegEx-based marker extraction achieved high performance metrics: Recall of 0.97, Precision of 0.84, and F1 score of 0.90.
  • Correlations: Statistically significant correlations were found between the presence of at least one comorbidity and the diagnosis of asthma, rhinosinusitis, nasal polyps, and bronchiectasis.

Limitations and Future Work

The authors acknowledge several limitations:

  • Information Omission: Despite accurate marker extraction, there's a possibility of omitting information relevant to the patient's overall health status, as the focus was on specific pathologies.
  • Pre-defined Pathologies: The selection of pathologies of interest before the marker extraction step might have led to overlooking other potentially useful information.
  • Text Formatting Errors: Errors in sentence formatting (e.g., lack of punctuation) or spelling mistakes could have influenced the marker extraction process, although RegEx aims to mitigate this.

Future work could involve:

  • Investigating Treatment Response: Exploring the correlation between identified patient clusters, their treatment, clinical response, and modifications in therapeutic approaches within a multidisciplinary ImmunoCenter compared to a standard allergy unit.
  • Broader Information Extraction: Expanding the NLP approach to capture a wider range of clinical information beyond predefined pathologies.
  • Addressing Negations: Further refining the NLP model to more accurately detect negations, which can impact precision.

Summary updated

Abstract

This analysis seems to be valid and is confirmed by the data from the literature. This means that NLP tools could have significant role in many other research fields of medicine, as it may help identify other important, and possibly previously neglected clusters of patients with comorbidities and commonalities. Another potential benefit of this approach lies in its potential ability to foster a multidisciplinary approach, using the same drugs to treat pathologies normally treated by physicians in different branches of medicine, thus saving resources and improving the pharmacological management of patients.
BibTeX
@inproceedings{morandini2022artificial,
  author = {Pierandrea Morandini and Maria Elena Laino and Giovanni Paoletti and Alessandro Carlucci and Tobia Tommasini and Giovanni Angelotti and J. Pepys and Giorgio Walter Canonica and Enrico Heffler and Victor Savevski and Francesca Puggioni},
  title = {{Artificial intelligence processing electronic health records to identify commonalities and comorbidities cluster at Immuno Center Humanitas}},
  booktitle = {Clinical},
  year = {2022},
  doi = {10.1002/clt2.12144},
}
Figure