Identification of routine blood derived hematological and lipid indices in ILD through machine learning; a retrospective case-control study

Li‐Chen Zhu, Yu Fu, Linchao Zhu, Yimin Yao, Li ChenFrontiers in Medicine 2025 · DOI

TL;DR

Researchers have developed a high-accuracy machine learning model that predicts the risk of Interstitial Lung Disease (ILD) using nothing more than routine blood work. By combining standard hematology counts with lipid profiles into novel "inflammatory-metabolic indices," the Random Forest-based model achieved an AUC of 0.885, offering a low-cost, non-invasive screening tool for primary care clinicians.

What Problem Does This Paper Address?

Interstitial Lung Disease (ILD) is a group of disorders characterized by lung inflammation and scarring (fibrosis). Early detection is critical for patient survival, yet the "gold standard" for diagnosis—High-Resolution Computed Tomography (HRCT) and pulmonary function tests—presents significant barriers. These methods are expensive, involve radiation exposure, and require specialized equipment often unavailable in primary care or resource-limited settings.

While specific biomarkers like KL-6 exist, they aren't part of standard check-ups. This paper addresses the need for a diagnostic "early warning system" built on data that is already collected in almost every clinical encounter: the complete blood count (CBC) and lipid panel.

Methods

The study utilized a retrospective case-control design, analyzing data from 891 patients. The researchers employed a rigorous "triple-filter" feature selection process to find the most predictive biomarkers, using three distinct algorithms:

  1. LassoCV (Least Absolute Shrinkage and Selection Operator)
  2. SVM-RFE-CV (Support Vector Machine-Recursive Feature Elimination)
  3. Boruta (a random forest-based feature wrapper)

After identifying the core features, they trained and compared eight different machine learning architectures, including XGBoost, LightGBM, and Logistic Regression. The Random Forest (RF) model emerged as the top performer.

A key methodological novelty was the calculation of derived indices, specifically the Neutrophil-to-HDL-C ratio (NHR) and Lymphocyte-to-HDL-C ratio (LHR), which bridge the gap between systemic inflammation and metabolic status.

Datasets

  • Primary Cohort: 603 patients (201 ILD cases, 402 age- and gender-matched controls) from the Hubin Campus of Zhejiang Provincial Hospital of Traditional Chinese Medicine.
  • External Validation Cohort: 288 patients from two additional medical centers to test the model’s generalizability.
  • Data Points: Clinical info, CBC data (leukocytes, differentials, etc.), and lipid metabolism indicators (TG, CHOL, HDL-C, LDL-C).

Key Insights and Ideas

  • The Power of Ratios: The study found that individual markers (like neutrophil count) are less predictive than ratios that combine inflammatory cells with metabolic markers like HDL-C. The NHR and LHR ratios reflect the "tug-of-war" between pro-inflammatory processes and the anti-inflammatory/antioxidant protection provided by high-density lipoproteins.
  • HDL-C as a Protective Signal: The results reinforce the idea that HDL-C isn't just for heart health; its anti-fibrotic properties make it a significant negative predictor for ILD.
  • Random Forest Wins on Robustness: While "boosted" models like XGBoost often win Kaggle competitions, the Random Forest model proved most stable here, maintaining high accuracy across internal and external datasets without overfitting.
  • Accessibility is the Goal: The authors didn't just publish a paper; they launched an online prediction tool. This allows a clinician to input six simple values from a standard lab report to get an immediate ILD risk estimate.

Key Results

  • Top Features: Six markers were identified as the "essential" set: Neutrophil %, Lymphocyte %, Monocyte %, Hemoglobin (HGB), NHR, and LHR.
  • Model Performance: The Random Forest model achieved:
  • AUC (Test Set): 0.885
    • AUC (External Validation): 0.849
    • Accuracy: ~86.8% in the test set.
  • Comparison: The model significantly outperformed individual biomarkers. For instance, while NHR alone had an AUC of 0.793, the integrated RF model pushed that performance much higher.

Limitations and Future Work

The authors acknowledge several hurdles:

  • Phenotype Bias: The cohort was heavily dominated by CTD-ILD (Connective Tissue Disease-related ILD), accounting for 87% of cases. The model's performance on other types, like Idiopathic Pulmonary Fibrosis (IPF) or sarcoidosis, remains to be fully validated.
  • Retrospective Nature: As a retrospective study, there is inherent risk of selection bias.
  • Sample Size: While nearly 900 patients is a solid start, the authors aim to validate the tool in much larger, multicenter prospective cohorts to ensure it works across diverse global populations.

Ultimately, this research moves us closer to a world where a routine physical's blood work can flag a serious lung condition years before it becomes a crisis.

Summary updated

Abstract

Based on these findings, we developed an online prediction tool to assist primary care clinicians in assessing the risk of ILD in suspected cases. Our results indicate that the random forest model exhibits high accuracy and clinical utility for early ILD prediction, providing a novel tool and methodology for early diagnosis and intervention. Future studies will focus on further optimizing the model and validating it in larger multicenter cohorts.
BibTeX
@inproceedings{zhu2025identification,
  author = {Li‐Chen Zhu and Yu Fu and Linchao Zhu and Yimin Yao and Li Chen},
  title = {{Identification of routine blood derived hematological and lipid indices in ILD through machine learning; a retrospective case-control study}},
  booktitle = {Frontiers in Medicine},
  year = {2025},
  doi = {10.3389/fmed.2025.1633713},
}
Figure