Prognostic Model for EGFR-Mutant NSCLC With Leptomeningeal Metastasis: LM-Index Development and External Validation

    This article presents a multicenter retrospective study that developed and externally validated LM-Index, a prognostic stratification model for patients with EGFR-mutant non-small cell lung cancer (NSCLC) and leptomeningeal metastasis (LM). Using a development cohort of 350 patients and an independent external validation cohort of 302 patients, the study applied Cox proportional hazards regression to identify five independent predictors of overall survival and converted their coefficients into an integer risk score and nomogram. Model performance was assessed through discrimination, calibration, decision curve analysis, and Kaplan-Meier survival analysis. The resulting tool is designed to support consistent risk assessment, 12- and 24-month overall survival prediction, and individualized management.

    Background and Clinical Challenges

    Leptomeningeal metastasis (LM) is one of the most severe central nervous system (CNS) complications in advanced non-small cell lung cancer (NSCLC). With the widespread use of targeted therapy, overall survival in patients with EGFR-mutant NSCLC has been significantly extended, but this has also increased opportunities for tumor cells to spread within the CNS, and the incidence of LM is rising. LM can cause severe headache, nausea and vomiting, cognitive decline, and a wide range of complex neurological symptoms, greatly impairing quality of life and worsening survival outcomes.

    In current clinical practice, physicians often rely on a combination of performance status scores, imaging findings, and cerebrospinal fluid (CSF) test results for empirical judgment. However, this approach lacks a quantitative and standardized multidimensional prognostic stratification tool, leading to poor consistency in risk assessment across medical centers. To address this gap, a rigorously validated risk prediction model is needed to support overall survival (OS) prediction and individualized management in precision medicine.

    Data Architecture and Cohort Design

    A multicenter retrospective study provided a solution to this clinical challenge. The study constructed a clinical dataset of 652 patients with EGFR-mutant NSCLC and LM. The dataset was divided according to standard machine learning and statistical modeling practice:

    • Model development cohort: 350 patients, used for feature selection, model training, and weight fitting.
    • Independent external validation cohort: 302 patients, used to evaluate the model’s generalization ability and stability on an unseen data distribution.

    The core prediction endpoint was overall survival (OS). This time-to-event modeling task typically requires survival analysis methods to handle censored data.

    Feature Engineering and LM-Index Construction

    The core of model construction was to extract features with strong predictive value from complex clinical indicators. The study team first collected baseline clinical characteristics, imaging findings, and CSF-related indicators in the development cohort, and then used Cox proportional hazards regression for analysis.

    Multivariable Cox Regression and Feature Selection

    Through univariate and multivariable Cox regression analyses, the researchers removed highly collinear variables and variables without statistical significance. Five predictors independently associated with poorer OS were ultimately identified.

    Integer Risk Score Design

    To make the model clinically interpretable and easy to use, the researchers converted the regression coefficients (Beta values) of each independent predictor into integer risk weights and constructed the LM-Index scoring system. A nomogram was also developed based on these features, allowing clinicians to directly estimate 12- and 24-month OS probabilities by adding simple scores.

    The following pseudocode shows the simplified scoring logic:

    1. def calculate_lm_index(patient_data):
    2. score = 0
    3. # 1. Performance status score
    4. if patient_data['ECOG_PS'] >= 3:
    5. score += weight_ecog
    6. # 2. Brain parenchymal metastasis
    7. if patient_data['brain_metastasis'] == True:
    8. score += weight_brain_met
    9. # 3. MRI findings
    10. if patient_data['MRI_leptomeningeal_enhancement'] == True:
    11. score += weight_mri
    12. # 4. Cerebrospinal fluid cytology
    13. if patient_data['CSFC'] == 'Positive':
    14. score += weight_csfc
    15. # 5. Intracranial pressure level
    16. if patient_data['ICP'] > 220:
    17. score += weight_icp
    18. return score, predict_os_probability(score)

    Key Predictors

    The five key features selected by the model characterize disease burden and physiological reserve from different dimensions:

    1. MRI evidence of leptomeningeal enhancement: Among all features, this had the strongest association with poorer OS, with a hazard ratio (HR) of 2.201 (95% CI: 1.585–3.057, P<0.001). This indicates that imaging-confirmed leptomeningeal involvement is the primary prognostic factor.
    2. ECOG PS ≥ 3: Performance status is a fundamental indicator of a patient’s ability to tolerate treatment. The HR was 2.044 (95% CI: 1.538–2.716, P<0.001), suggesting that patients with poor performance status face a very high risk of death and often cannot benefit from aggressive antitumor therapy.
    3. Concomitant brain parenchymal metastasis: The HR was 1.874 (95% CI: 1.369–2.566, P<0.001). Dual involvement of the brain parenchyma and leptomeninges reflects extensive tumor dissemination within the CNS.
    4. Positive cerebrospinal fluid cytology (CSFC): The HR was 1.824 (95% CI: 1.243–2.678, P=0.002). Detection of malignant cells in CSF is the gold standard for LM diagnosis, and a positive result directly confirms tumor cell activity in the CSF circulation.
    5. Intracranial pressure (ICP) > 220 mmH₂O: The HR was 1.784 (95% CI: 1.353–2.352, P<0.001). Intracranial hypertension not only causes severe symptoms but also limits the penetration of systemic drugs and affects treatment efficacy.

    Model Evaluation and External Validation

    A reliable prediction model must undergo rigorous performance validation. The study comprehensively evaluated LM-Index from three dimensions: discrimination, calibration, and clinical net benefit:

    • Discrimination assessment: The concordance index was used to evaluate the model’s ability to distinguish high- and low-risk patients and to ensure consistency between predicted probabilities and the ordering of actual survival outcomes.
    • Calibration assessment: Calibration curves were used to compare the model-predicted 12- and 24-month OS probabilities with observed survival rates, validating the accuracy of absolute probability predictions.
    • Decision curve analysis (DCA): DCA was used to evaluate the clinical net benefit of the model across different threshold probabilities, demonstrating that LM-Index can provide positive benefit in actual decision-making.
    • External validation and survival analysis: In the independent external validation cohort of 302 patients, the model showed excellent stability. Kaplan-Meier survival analysis comparing different risk groups showed significant statistical stratification in OS between the high-risk and low-risk groups.

    Summary and Clinical Translation Value

    The LM-Index risk prediction model successfully constructs a survival risk stratification tool for patients with EGFR-mutant NSCLC and LM by integrating five routinely available clinical indicators. From data collection and feature engineering to statistical modeling and external validation, the model forms a complete technical loop. Its clinical application can effectively improve the consistency and accuracy of prognostic judgment, provide strong data support for physicians in developing individualized treatment strategies and planning key points for physician-patient communication, and offer a valuable engineering practice reference for the development of future medical time-to-event prediction models.