How to Build a Prognostic Risk Stratification Model for EGFR-Mutant NSCLC With Leptomeningeal Metastasis
This article explains how a prognostic risk-stratification model was developed for patients with EGFR-mutant non-small cell lung cancer (NSCLC) and leptomeningeal metastasis (LM). Based on 652 patients, the study identified five independent adverse prognostic factors—MRI evidence of leptomeningeal enhancement, ECOG performance status (PS) ≥3, parenchymal brain metastases, positive cerebrospinal fluid cytology, and intracranial pressure >220 mmH₂O—and translated them into an integer-based LM-Index and a nomogram. The model was evaluated in development and external validation cohorts, with reported discrimination and calibration, and divided patients into low-, intermediate-, and high-risk groups with significantly different survival curves. The article covers the clinical rationale, study design, factor selection, score construction, validation approach, clinical interpretation, limitations, and FAQs.
Key Takeaways
This article explains how to build a prognostic risk-stratification model for patients with EGFR-mutant NSCLC and leptomeningeal metastasis (LM).
The study included 652 patients with EGFR-mutant NSCLC and LM. It identified five independent adverse prognostic factors:
- MRI evidence of leptomeningeal enhancement
- ECOG PS ≥3
- Parenchymal brain metastases
- Positive cerebrospinal fluid cytology
- Intracranial pressure >220 mmH₂O
These factors were used to construct an integer-based LM-Index score and a nomogram. The model was evaluated in a development cohort and an external validation cohort. It divided patients into low-risk, intermediate-risk, and high-risk groups, with clearly separated survival curves.
The article follows this structure: clinical problem → study design → factor selection → score construction → performance validation → clinical interpretation → limitations.
1. Why Do Prognoses in LM Patients Need Quantitative Risk Stratification?
Leptomeningeal metastasis (LM) is a serious neurological complication of advanced non-small cell lung cancer (NSCLC). Tumor cells spread through the cerebrospinal fluid and implant in the meninges and subarachnoid space. Patients may present with headache, vomiting, cognitive decline, cranial nerve palsies, and other neurological symptoms. These manifestations can substantially impair quality of life and survival.
For patients with EGFR-driven NSCLC, advances in targeted therapy and systemic treatment have prolonged control of systemic disease. As a result, the central nervous system—especially the leptomeninges—has become a more common site of recurrence or progression. This has increased both the detection rate and clinical attention paid to LM.
The prognosis of LM patients is highly heterogeneous. Some patients can achieve relatively long survival with timely intervention, while others deteriorate rapidly. In clinical practice, physicians often rely on empirical judgment based on performance status, imaging findings, and cerebrospinal fluid results. However, there is no unified standard, and consistency between physicians and centers is limited.
Therefore, an externally validated, quantitative, and reproducible multidimensional prognostic stratification tool is needed to improve consistency in risk assessment and support treatment decision-making.
2. How Was the Study Designed and How Were the Cohorts Divided?
The study used a multicenter retrospective design. The prediction endpoint was overall survival (OS). The authors developed and validated an LM-Index model for patients with EGFR-mutant NSCLC and LM.
A total of 652 patients were included:
- 350 patients formed the model development cohort, used for factor selection and model construction.
- 302 patients formed an independent external validation cohort, used to assess model stability and generalizability.
Separating the validation cohort from the development cohort is an important design strategy for reducing overfitting and testing real-world performance. The development cohort addresses whether the model can fit known data. The external validation cohort addresses whether the model remains reliable on new samples. Together, they support the credibility of the findings.
A retrospective design is practical for accumulating samples and performing multivariable analysis. However, it places higher demands on data-source consistency and standardized variable collection.
3. How Were Independent Prognostic Factors Selected?
In the development cohort, the researchers first performed univariate and multivariable Cox proportional hazards regression to identify factors independently associated with OS. Candidate variables included clinical, imaging, and cerebrospinal fluid indicators.
Multivariable Cox regression ultimately identified five factors independently associated with worse OS:
| Predictor | HR | 95% CI | P value | Domain |
|---|---|---|---|---|
| MRI evidence of leptomeningeal enhancement | 2.201 | 1.585–3.057 | <0.001 | Imaging |
| ECOG PS ≥3 | 2.044 | 1.538–2.716 | <0.001 | Patient status |
| Parenchymal brain metastases | 1.874 | 1.369–2.566 | <0.001 | Imaging |
| Positive CSF cytology | 1.824 | 1.243–2.678 | 0.002 | CSF |
| Intracranial pressure >220 mmH₂O | 1.784 | 1.353–2.352 | <0.001 | CSF |
The sample sizes, HRs, and 95% CIs shown above are from the original study material. When formally citing this work, the original publication should be referenced. If no definitive source is available, these values should be treated as methodological illustrative values.
These five factors cover three domains: patient status, imaging, and cerebrospinal fluid findings. This combination balances clinical accessibility with prognostic discrimination. All five factors have HR values greater than 1, and their confidence intervals do not cross 1, indicating that each represents an independent adverse prognostic signal.
Technical note: Intracranial pressure should ideally be measured as the opening pressure during lumbar puncture and recorded in mmH₂O. During model development, the proportional hazards assumption should be checked, for example using Schoenfeld residual tests. If the assumption is not met, time-dependent covariates or stratified analysis should be considered.
4. How Was the Model Turned into a Computable Score and Nomogram?
The researchers used the regression coefficients from the Cox model to construct an integer-based score. In Cox regression, the coefficient β can be derived from the hazard ratio as β = ln(HR). Each factor was converted into an integer score, and a nomogram was also created.
The illustrative score assignment is shown below. This table is provided for methodological illustration only.
| Factor | Score When Positive or Threshold Is Met |
|---|---|
| MRI evidence of leptomeningeal enhancement | 3 |
| ECOG PS ≥3 | 3 |
| Parenchymal brain metastases | 2 |
| Positive CSF cytology | 2 |
| Intracranial pressure >220 mmH₂O | 2 |
The total score ranges from 0 to 12. The illustrative risk-group cutoffs are:
- Low risk: 0–2 points
- Intermediate risk: 3–5 points
- High risk: 6–12 points
Example calculation, for illustration only:
A patient has MRI evidence of leptomeningeal enhancement (+3), ECOG PS = 2 (0 points), parenchymal brain metastases (+2), positive CSF cytology (+2), and normal intracranial pressure (0 points).
Total score = 3 + 0 + 2 + 2 + 0 = 7 points.
This patient would fall into the high-risk group, corresponding to a lower estimated 12-month OS probability.
The core idea of the nomogram is to convert a multivariable model into a graphical tool. Each predictor is mapped to a scale, the corresponding points are read off, and the points are summed. The total score can then be used to estimate 12-month and 24-month OS probabilities.
Compared with presenting a regression equation directly, a nomogram transforms abstract coefficients into a more intuitive tool. It can help clinicians perform individualized assessments quickly in outpatient or bedside settings and improve the communication of prognostic information between physicians and patients.
5. How Was Model Performance Evaluated?
A reliable prognostic model should be assessed for discrimination, calibration, and clinical utility. In this study, the model was evaluated in both the development cohort and the external validation cohort.
- Discrimination, measured by the C-index: A C-index closer to 1 indicates stronger ability to distinguish high-risk from low-risk patients. The illustrative C-index was approximately 0.75 in the development cohort and approximately 0.72 in the external validation cohort, suggesting that the model may be generalizable.
- Calibration, assessed by calibration curves: Calibration compares predicted probabilities with observed event rates. For example, among patients predicted to have a 30% risk, approximately 30% should experience the event. The illustrative calibration curves in both cohorts were close to the diagonal line, with no obvious systematic overestimation or underestimation.
- Clinical utility, assessed by decision curve analysis (DCA): Across most clinically relevant risk thresholds, using the model provided net benefit compared with treating all patients or treating none.
- Stratification ability, assessed by Kaplan–Meier analysis and between-group comparisons: The illustrative median OS was approximately 20 months in the low-risk group, 11 months in the intermediate-risk group, and 5 months in the high-risk group. The survival curves of the three groups were significantly separated, with log-rank P <0.001.
The C-index values, calibration results, median OS values, and log-rank P value above are illustrative. They are used only to explain how validation methods and results can be presented. The exact values should follow the original research.
Applying all four types of evaluation to both the development and external validation cohorts is essential to determine whether a model is not only able to fit the development data but also able to generalize to new patients. The significant separation of survival curves among risk groups in the external validation cohort is key evidence that the model has real prognostic stratification value.
6. What Do Different Risk Groups Mean for Treatment Decisions?
The LM-Index integrates dispersed clinical information into a quantitative risk-stratification result. It is intended to assist—not replace—treatment decision-making. The following points are general methodological considerations and must be adapted to individual patient circumstances.
- Low-risk group: Patients generally have better overall condition and relatively lower disease burden. They may be more likely to tolerate and benefit from systemic therapy, including sensitive EGFR-TKI-based regimens or combination approaches, together with local therapy. Active disease control may be pursued.
- Intermediate-risk group: Treatment often requires a combination of systemic therapy and local or intrathecal therapy. Attention should be given to both intracranial disease control and symptom management, with close follow-up of cerebrospinal fluid findings and imaging.
- High-risk group: Patients often have poorer performance status and higher intracranial disease burden. Treatment intensity must be balanced against supportive care needs. Intracranial pressure management, symptom relief, and shared decision-making are particularly important.
For clinical research, risk stratification can help enrich or balance patients with different prognoses and improve study efficiency. For physician–patient communication, quantified survival probabilities can make prognostic discussions more transparent and predictable.
Medical disclaimer: The LM-Index is intended for research and clinical decision support. It cannot replace a physician’s clinical judgment. The risk groups and treatment considerations described above are methodological examples only. They require prospective validation before clinical implementation. Actual diagnosis and treatment should follow the advice of the treating physician.
7. What Are the Limitations and Future Directions?
As a retrospective study, its conclusions may be affected by the completeness of previous medical records, differences in measurement practices across centers, and potential confounders that were not included. Subjective variables, such as performance status assessment and the timing of cerebrospinal fluid collection, may also introduce bias.
Future work could use prospective, larger-scale multicenter data for further validation. The model could also incorporate molecular subtypes, the number of prior targeted therapy lines, and dynamic cerebrospinal fluid monitoring to improve accuracy and applicability.
For long-term reliable use in real-world settings, the model needs periodic recalibration and revalidation.
Frequently Asked Questions (FAQ)
What is the LM-Index?
The LM-Index is a prognostic risk score and nomogram for patients with LM. It integrates five commonly available clinical factors and outputs a low-, intermediate-, or high-risk classification, along with estimated 12-month and 24-month OS probabilities.Who is the model intended for?
It is intended for patients with EGFR-mutant NSCLC and leptomeningeal metastasis, based on confirmed diagnosis or clinical assessment. It is intended for research stratification and clinical decision support.How is the risk score calculated?
Points are assigned according to MRI leptomeningeal enhancement, ECOG PS ≥3, parenchymal brain metastases, positive CSF cytology, and intracranial pressure >220 mmH₂O. The points are summed, and the total score is compared with the predefined cutoffs to determine the risk group.What are common pitfalls?
Common mistakes include treating model probabilities as an individual’s inevitable outcome, ignoring the role of performance status and supportive care, and directly applying a score without local calibration.
