A new classifier constructed with platelet features for malignant and benign pulmonary nodules based on prospective real-world data

Ruiling Zu, Lin Wu, Rong Zhou, Xiaoxia Wen, Bangrong Cao, Shan Liu, Guishu Yang, Ping Leng, Yan Li, Li Zhang, Xiao-Yu Song, Yao Deng, Kaijiong Zhang, Chang Liu, Yuping Li, Jian Huang, Dongsheng Wang, Guiquan Zhu, Huaichao LuoJournal of Cancer 2022 · DOI

TL;DR

This paper introduces the SCHC model, a novel classifier that uses platelet features, age, and nodule size to distinguish between malignant and benign pulmonary nodules. By leveraging readily available blood test data, the SCHC model offers a promising, low-cost approach to improve early-stage lung cancer diagnosis and management, potentially reducing unnecessary invasive procedures.

What Problem Does This Paper Address?

Lung cancer remains a leading cause of cancer death globally. Low-dose computed tomography (LDCT) is a crucial screening tool for detecting pulmonary nodules, but differentiating between benign and malignant nodules based on imaging alone is challenging. This ambiguity often leads to unnecessary invasive biopsies or delayed diagnosis, impacting patient outcomes. Existing clinical models, while helpful, still struggle with accurate discrimination, highlighting the need for more effective and accessible diagnostic tools.

Methods

The researchers developed a predictive model using eXtreme Gradient Boosting (XGBoost), a powerful machine learning algorithm known for its accuracy in classification tasks. They integrated computed tomographic (CT) data, clinical information, and, crucially, novel platelet-derived features from both whole blood and platelet-rich plasma (PRP) samples. The model was trained on a development cohort and then validated on an external cohort. Performance was evaluated using standard metrics like Receiver Operating Characteristic (ROC) curves, Area Under the Curve (AUC), Net Reclassification Improvement (NRI), Integrated Discrimination Improvement (IDI), and Net Benefit (NB). They also developed a user-friendly web tool to make the SCHC model accessible to clinicians.

Datasets

The study utilized data from two hospitals in China: Sichuan Cancer Hospital and Sichuan Provincial People's Hospital.

  • Development Cohort: 419 participants, collected between January 2018 and December 2020. This cohort was further split into a training set (70%) and a testing set (30%).
  • External Validation Cohort: 62 participants, collected between March 2021 and April 2021.

Key Insights and Ideas

  • Platelets as Biomarkers: The study strongly suggests that platelet features, easily obtainable from routine blood tests, hold significant potential for diagnosing lung cancer. This moves beyond traditional imaging and clinical data.
  • XGBoost's Power: The XGBoost algorithm proved effective in identifying the most important features for discrimination, allowing for the construction of a parsimonious yet powerful model.
  • Synergy of Features: The SCHC model's success lies in its combination of readily available clinical data (age), imaging-derived data (nodule size), and novel hematological data (platelet counts and plateletcrit).
  • Clinical Utility: The development of a web tool integrating the SCHC model with existing models (VA, MC, BU) makes this advanced diagnostic capability accessible and practical for clinical use.

Key Results

The developed SCHC model, incorporating age, platelet counts in PRP (pPLT), plateletcrit in PRP (pPCT), nodule size, and plateletcrit in whole blood (bPCT), demonstrated strong performance in the development cohort.

  • Development Cohort Performance: The SCHC model achieved good performance with AUCs of 0.76 (training) and 0.72 (testing).
  • Improved Reclassification: Compared to existing models (VA, MC, BU), the SCHC model showed significant improvements in reclassification as assessed by NRI and IDI, indicating better discrimination.
  • Net Benefit: Decision curve analysis showed that the SCHC model provided the highest net benefit across various probability thresholds, suggesting greater clinical value.
  • External Validation: The model's performance was less robust in the external validation cohort (AUC=0.52), which the authors attribute to potential variations in sample processing and a smaller validation dataset.

Limitations and Future Work

  • External Validation Performance: The SCHC model's performance was notably lower in the external validation cohort. The authors suggest this could be due to variations in PRP sample processing (centrifugation, laboratory environment, manipulator technique) and the inherent activation of platelets.
  • Small Validation Set: The limited size of the external validation cohort may have contributed to the reduced performance.
  • Future Directions: The authors plan to increase the validation sample size and develop a more standardized operating manual to minimize variability in platelet feature measurements. They also aim to further refine the model and its integration into clinical practice. The study also highlights the broader potential of common laboratory results for diagnosing various diseases, not just cancers.

Summary updated

Abstract

<b>Objectives:</b> As the pulmonary nodules were hard to be discriminated as benignancy or malignancy only based on imageology, a prospective and observational real-world research was devoted to develop and validate a predictive model for managing the diagnostic challenge. <b>Methods:</b> This study started in 2018, and a predictive model was constructed using eXtreme Gradient Boosting (XGBoost) based on computed tomographic, clinical, and platelet data of all the eligible patients. And the model was evaluated and compared with other common models using ROC curves, continuous net reclassification improvement (NRI), integrated discrimination improvement (IDI), and net benefit (NB). Subsequently, the model was validated in an external cohort. <b>Results:</b> The development group included 419 participants, while there were 62 participants in the external validation cohort. The most accurate XGBoost model called SCHC model including age, platelet counts in platelet rich plasma samples (pPLT), plateletcrit in platelet rich plasma samples (pPCT), nodule size, and plateletcrit in whole blood samples (bPCT). In the development group, the SCHC model performed well in whole group and subgroups. Compared with VA, MC, BU model, the SCHC model had a significant improvement in reclassification as assessed by the NRI and IDI, and could bring the patients more benefits. For the external validation, the model performed not as well. The algorithm of SCHC, VA, MC, and BU model were first integrated using a web tool (http://i.uestc.edu.cn/SCHC). <b>Conclusions:</b> In this study, a platelet feature-based model could facilitate the discrimination of early-stage malignancy from benignancy patients, to ensure accurate diagnosis and optimal management. This research also indicated that common laboratory results also had the potential in diagnosing cancers.
BibTeX
@article{zu2022new,
  author = {Ruiling Zu and Lin Wu and Rong Zhou and Xiaoxia Wen and Bangrong Cao and Shan Liu and Guishu Yang and Ping Leng and Yan Li and Li Zhang and Xiao-Yu Song and Yao Deng and Kaijiong Zhang and Chang Liu and Yuping Li and Jian Huang and Dongsheng Wang and Guiquan Zhu and Huaichao Luo},
  title = {{A new classifier constructed with platelet features for malignant and benign pulmonary nodules based on prospective real-world data}},
  journal = {Journal of Cancer},
  year = {2022},
  doi = {10.7150/jca.67428},
}
Figure