Scholarlink

Background <p>Rapid screening of hematolymphoid malignancies (HMs) by complete blood cell count (CBC) is crucial for choosing the next appropriate workup and initiating timely treatment for critical cases such as acute promyelocytic leukemia. Commonly presented as leukocytosis and abnormal differential, HMs could be misdiagnosed with diverse etiologies including reactive status of bacterial or viral infection, which also require early diagnosis and treatment to prevent severity, such as sepsis. Conventional workflow based on morphology review for HM screening remains time-consuming, labor-intensive, expertise-dependent and subjective. This study aimed to develop and validate an interpretable and cost-effective artificial intelligence (AI)-assisted model, to enable rapid and accurate prediction of HMs mixed with acute bacterial and viral infection, thus informing proper further workup and timely treatment.</p> Methods <p>We retrospectively collected a multicenter dataset containing 4996 CBC records from patients presented with leukocytosis and finally diagnosed of either diseases from six prevalent HMs and acute bacterial and/or acute viral infection for model development and validation. This eight-nomial classification model, on a multimodal basis, integrates the digit parameters and its matching WBC differential scattergram from a single CBC test via a convolutional fusion deep neural network, and was compared with the digit- and scattergram-based unimodal model, as well as the current CBC review rules for performance evaluation.</p> Results <p>The multimodal mDNN-cHM outperformed either unimodal model, as is demonstrated by metrics including precision, recall, F1-score and accuracy (all <i>p</i> &lt; 0.05). The model showed striking diagnostic accuracy in both internal and external validations (the area under the ROC curve [AUC-ROC] all above 0.95, ranging from 0.98–1.00 internally and 0.95–1.00 externally). Further, mDNN-cHM demonstrated a higher prediction rate than the flag prompt by the current CBC review rules.</p> Conclusions <p>Our study presents a proof-of-concept AI classifier for HMs in a real-time manner, which potentially maximizes the diagnostic value of the low-cost and widely-accessible CBC within 0.5 h after sample receipt. The framework also underscores the potential of interpretable AI in accelerating early disease diagnosis, mitigating human error and enlightening clinical decision-making via CBC, which complements and enhances standard-of-care diagnostics, particularly for urgent and high-risk diseases including leukemia/lymphoma and acute infection.</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

Development and validation of a novel multimodal deep neural network model based on CBC digit parameters and scattergrams for rapid hematolymphoid malignancy classification: a multicenter cohort study

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月8日

查看全文
  • Hongyan Liao,
  • Siyuan Chen,
  • Qiang Meng,
  • Mingxia Tang,
  • Shuyi Hou,
  • Jie Zhao,
  • Yan Liu,
  • Qin Zheng

摘要

Background

Rapid screening of hematolymphoid malignancies (HMs) by complete blood cell count (CBC) is crucial for choosing the next appropriate workup and initiating timely treatment for critical cases such as acute promyelocytic leukemia. Commonly presented as leukocytosis and abnormal differential, HMs could be misdiagnosed with diverse etiologies including reactive status of bacterial or viral infection, which also require early diagnosis and treatment to prevent severity, such as sepsis. Conventional workflow based on morphology review for HM screening remains time-consuming, labor-intensive, expertise-dependent and subjective. This study aimed to develop and validate an interpretable and cost-effective artificial intelligence (AI)-assisted model, to enable rapid and accurate prediction of HMs mixed with acute bacterial and viral infection, thus informing proper further workup and timely treatment.

Methods

We retrospectively collected a multicenter dataset containing 4996 CBC records from patients presented with leukocytosis and finally diagnosed of either diseases from six prevalent HMs and acute bacterial and/or acute viral infection for model development and validation. This eight-nomial classification model, on a multimodal basis, integrates the digit parameters and its matching WBC differential scattergram from a single CBC test via a convolutional fusion deep neural network, and was compared with the digit- and scattergram-based unimodal model, as well as the current CBC review rules for performance evaluation.

Results

The multimodal mDNN-cHM outperformed either unimodal model, as is demonstrated by metrics including precision, recall, F1-score and accuracy (all p < 0.05). The model showed striking diagnostic accuracy in both internal and external validations (the area under the ROC curve [AUC-ROC] all above 0.95, ranging from 0.98–1.00 internally and 0.95–1.00 externally). Further, mDNN-cHM demonstrated a higher prediction rate than the flag prompt by the current CBC review rules.

Conclusions

Our study presents a proof-of-concept AI classifier for HMs in a real-time manner, which potentially maximizes the diagnostic value of the low-cost and widely-accessible CBC within 0.5 h after sample receipt. The framework also underscores the potential of interpretable AI in accelerating early disease diagnosis, mitigating human error and enlightening clinical decision-making via CBC, which complements and enhances standard-of-care diagnostics, particularly for urgent and high-risk diseases including leukemia/lymphoma and acute infection.