Scholarlink

<p>Cardiac amyloidosis (CA), a fatal and progressive cardiomyopathy is characterized by amyloid deposition within the myocardium. The main forms of CA include transthyretin amyloidosis (ATTR-CM; distinguished into a hereditary and wildtype form [ATTRv and ATTRwt]) and light-chain amyloidosis (AL). Transthyretin amyloidosis with cardiomyopathy (ATTR-CM) is increasingly recognized among heart failure (HF) patients, despite its underestimated prevalence as a rare disease. ATTR-CM is challenging to diagnose due to its broad phenotypic overlap with HF, yet a timely differentiation of CA is crucial, as effective treatment can significantly improve quality of life. Machine learning (ML) approaches could enable earlier detection of ATTR-CM in routine care data. Data of ATTR-CM patients from the Amyloidosis Center Charité Berlin (2002-2023) and data from a cohort with heart failure and aortic stenosis screened for ATTR-CM were mapped from ICD-10-GM to ICD-10-CM and evaluated with an open-source random forest (RF) model based on ICD-10-CM codes which was previously validated in multiple external cohorts in the US and UK. In the cohort of patients with HF and aortic stenosis, cases of ATTR-CM were recognized using the algorithm, although it was not possible to distinguish between ATTRv-CM and ATTRwt-CM. Sensitivity and specificity were significantly lower than in the external validations in the US and UK. Model performance depended strongly on coding granularity and feature availability and contributed to the diminished predictive power. While the RF model showed moderate transferability under enriched conditions, reliance of geographically specific ICD-10 systems limits broader applicability due to coding discrepancies and information loss. These findings underscore the need for improved semantic harmonization of routine data, as well as the integration of rare disease specific ontologies such as the Human Phenotype Ontology (HPO) and ORPHAcodes to enhance cross-border ML-based assistance for timely diagnosing patients with rare diseases. </p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Transferability of a US claims-based machine learning model for ATTRwt-CM identification: a retrospective evaluation in a German setting

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月24日

查看全文
  • Harisa Muratovic-Colic,
  • Miriam R. Hübner,
  • Jan-Filip Rehburg,
  • Isabel Mattig,
  • Paul J. Wetzel,
  • Katrin Wrede-Wihl,
  • Helena F. Pernice,
  • Gina Barzen,
  • Nicolas Wieder,
  • Jakub Piwowarski,
  • Stephan Bohl,
  • Vera von Landenberg-Roberg,
  • Daniel Messroghli,
  • Sebastian Spethmann,
  • Richard Röttger,
  • Josef Schepers,
  • Katrin Hahn

摘要

Cardiac amyloidosis (CA), a fatal and progressive cardiomyopathy is characterized by amyloid deposition within the myocardium. The main forms of CA include transthyretin amyloidosis (ATTR-CM; distinguished into a hereditary and wildtype form [ATTRv and ATTRwt]) and light-chain amyloidosis (AL). Transthyretin amyloidosis with cardiomyopathy (ATTR-CM) is increasingly recognized among heart failure (HF) patients, despite its underestimated prevalence as a rare disease. ATTR-CM is challenging to diagnose due to its broad phenotypic overlap with HF, yet a timely differentiation of CA is crucial, as effective treatment can significantly improve quality of life. Machine learning (ML) approaches could enable earlier detection of ATTR-CM in routine care data. Data of ATTR-CM patients from the Amyloidosis Center Charité Berlin (2002-2023) and data from a cohort with heart failure and aortic stenosis screened for ATTR-CM were mapped from ICD-10-GM to ICD-10-CM and evaluated with an open-source random forest (RF) model based on ICD-10-CM codes which was previously validated in multiple external cohorts in the US and UK. In the cohort of patients with HF and aortic stenosis, cases of ATTR-CM were recognized using the algorithm, although it was not possible to distinguish between ATTRv-CM and ATTRwt-CM. Sensitivity and specificity were significantly lower than in the external validations in the US and UK. Model performance depended strongly on coding granularity and feature availability and contributed to the diminished predictive power. While the RF model showed moderate transferability under enriched conditions, reliance of geographically specific ICD-10 systems limits broader applicability due to coding discrepancies and information loss. These findings underscore the need for improved semantic harmonization of routine data, as well as the integration of rare disease specific ontologies such as the Human Phenotype Ontology (HPO) and ORPHAcodes to enhance cross-border ML-based assistance for timely diagnosing patients with rare diseases.