Real-world EHR-derived progression-free survival across successive lines of therapy informs metastatic breast cancer risk stratification

This paper presents a scalable framework that reconstructs metastatic breast cancer treatment lines from real-world EHR data using radiology-anchored progression evidence to generate individualized, interpretable progression-free survival estimates that enable robust risk stratification across diverse patient subgroups and successive lines of therapy.

The Gist

Imagine you are navigating a complex, ever-changing maze. This maze represents metastatic breast cancer. The goal isn't to "win" the maze in the traditional sense, but to keep moving through it for as long as possible without hitting a dead end (which, in medical terms, is the cancer growing or spreading).

Doctors have been trying to predict how long a patient can stay on a specific path (a treatment) before hitting a dead end. This is called Progression-Free Survival (PFS).

Here is the problem: The first path (first-line treatment) is well-mapped. But once patients have to switch to a second, third, or fourth path, the map gets blurry. Every patient is different, the treatments are highly customized, and the "dead ends" aren't always recorded clearly in the hospital's computer system. It's like trying to predict the weather in a city where everyone uses a different umbrella and the rain sensors are broken.

This paper introduces a smart, AI-powered navigation system built from real-world data to help predict how long a patient can stay on their current path before needing to switch.

The Big Idea: Building a "Time Machine" from Hospital Records

The researchers took a massive pile of messy, real-world hospital records (Electronic Health Records, or EHRs) from nearly 3,000 patients. They didn't just look at the numbers; they used a clever trick to clean up the data:

1. The "Anchor" Trick: Instead of guessing when the cancer grew based on a doctor's vague note, they used radiology reports (CT scans and MRIs) as the "anchor." They built a system that says, "Okay, the treatment started here, and the next scan clearly shows the cancer grew there. That's the end of this path." This created a clean, reliable timeline for every patient.
2. The "Snapshot" Approach: Imagine taking a photo of a patient the moment they start a new treatment. This photo includes their age, their tumor markers (chemical signals in the blood), their genetic makeup, and what the latest scan shows. The AI learns to look at this single snapshot and predict: "Based on this photo, how long will this person stay on this path?"

How the AI Learned (The "Student" Analogy)

Think of the AI model as a medical student who has read thousands of case files.

• The Training: The student was shown thousands of "snapshots" (start of treatment) and the "outcome" (how long the treatment actually worked).
• The Challenge: The student had to learn without cheating. They couldn't look at the future (e.g., they couldn't see that a patient switched drugs after the treatment started because it wasn't working). They had to make their prediction based only on what was known at the start.
• The Result: The student (a specific type of AI called a Gradient Boosted Survival model) became very good at guessing. It could look at a new patient's snapshot and say, "This patient is in the 'High Risk' group; they might need to switch treatments in 3 months," or "This patient is in the 'Low Risk' group; they might stay on this path for a year."

Why This is a Game-Changer

1. It Works for Everyone, Not Just the "Average" Patient
In the past, doctors often relied on data from clinical trials, which usually only include very specific types of patients. This AI learned from the "real world," where patients are messy, have different histories, and take different drugs. It's like learning to drive on a chaotic city street rather than just a quiet test track. The result? It works well for almost every type of breast cancer, whether it's aggressive or slower-growing.

2. It Explains Why (The "Flashlight" Effect)
Many AI models are "black boxes"—they give an answer but won't tell you why. This model is different. It acts like a flashlight, shining on the specific reasons for its prediction.

• Example: If the AI predicts a high risk, it might say, "It's not because of the patient's age, but because the latest scan shows a new spot in the liver, and their blood markers are rising fast."
• This helps doctors trust the prediction and understand what to watch out for.

3. It Helps Plan the Next Move
If the AI predicts a patient is "High Risk," the doctor knows to be proactive. They might schedule more frequent scans or start planning the next treatment now, rather than waiting until the current one fails. If the prediction is "Low Risk," the doctor might say, "Great, let's stick with this plan and not over-treat."

The Catch (Limitations)

The authors are honest about the flaws.

• The First Line is Tricky: The AI is slightly less accurate for the very first treatment a patient gets after the cancer spreads. It's like the AI is great at predicting the middle of a journey but needs more data to predict the very first step.
• Data Quality: The AI is only as good as the hospital records. If a doctor forgets to write down a scan result, the AI has to guess.

The Bottom Line

This paper is about building a smart, data-driven compass for doctors treating metastatic breast cancer. By cleaning up messy hospital data and training an AI to look at the "snapshot" of a patient's health at the start of a treatment, they can now give much better estimates of how long a treatment will last.

Instead of guessing in the dark, doctors can now say, "Based on the patterns of thousands of others, here is the most likely path for you, and here is exactly what we should watch for." It turns a chaotic, individual battle into a more predictable, manageable journey.

Technical Summary

1. Problem Statement

Metastatic breast cancer (mBC) is treated with successive lines of therapy (mLoTs). While Progression-Free Survival (PFS) is the standard metric for treatment efficacy, reliable real-world evidence for second-line and beyond (mLoT2+) is scarce.

• Challenges:
  • Data Fragmentation: Real-world Electronic Health Records (EHR) lack structured endpoints for disease progression; progression is often inferred from unstructured radiology reports.
  • Heterogeneity: Treatment regimens become highly individualized in later lines, making it difficult to generalize efficacy.
  • Confounding: Treatment selection is tightly coupled with evolving disease burden, creating confounding factors that standard trial designs cannot capture.
  • Data Leakage: Predictive models often inadvertently use "future" information (e.g., treatment switches due to early toxicity or progression) to predict outcomes, leading to inflated performance that is not actionable at the time of treatment initiation.

2. Methodology

The authors propose a framework to reconstruct clinically coherent mLoTs and train interpretable machine learning models to predict PFS at the start of each therapy line.

A. Cohort Construction & mLoT Segmentation

• Data Source: MSK-CHORD dataset (Memorial Sloan Kettering Cancer Center), comprising 2,881 patients with 8,791 metastatic lines of therapy.
• Segmentation Algorithm: Uses heuristic rules based on clinical conventions to define line boundaries:
  • A new line starts after a radiology-confirmed progression event (occurring $\ge$28 days after line start).
  • A new line starts if a new systemic drug is initiated $\ge$1 year after the current line start.
  • Local treatments (radiation, surgery) do not trigger a new line.
• PFS Labeling: Defined as time from line start to the first radiology-confirmed progression or death.
  • Anchoring: Progression is anchored to NLP-derived radiology assessments to ensure structured labeling.
  • Censoring: Administrative censoring at 2 years to stabilize weights.
• Inclusion Criteria: Strict filters applied to ensure label fidelity (e.g., requiring radiology assessments within 90 days prior to line start, excluding mixed chemo/endocrine regimens in a single line).

B. Feature Engineering

Features are extracted from five modalities at the line-start snapshot:

1. Clinicopathologic: Demographics, ECOG status.
2. Radiology: Tumor presence in specific organs (derived from reports within 90 days pre-line).
3. Tumor Markers: Kinetics of CA15–3 and CEA.
4. Planned Regimen: Drug agents intended for the current line (top 5 + "other").
5. Genomics: Somatic mutations (e.g., TP53) and copy-number aberrations.

Critical Design Choices to Prevent Leakage:

• Treatment Masking: Drugs initiated >28 days after line start are masked. This prevents the model from learning from post-initiation adaptations (e.g., switching drugs due to early progression), ensuring the prediction represents a baseline prognosis.
• Surveillance Suppression: Fine-grained surveillance patterns (measurement counts/timestamps) are omitted to prevent the model from using "care intensity" as a proxy for disease severity.

C. Model Selection & Training

• Models Benchmarked: CoxPH, DeepSurv, DeepHit, Random Survival Forest (RSF), and Gradient-Boosted Survival Analysis (GBSA).
• Validation: Nested Cross-Validation (5 outer folds, 3 inner folds) to prevent overfitting and ensure unbiased hyperparameter tuning.
• Selected Model: GBSA (Gradient-Boosted Survival Analysis with CoxPH loss) was selected for its superior discrimination (AUC) and calibration (Integrated Brier Score).

3. Key Results

A. Model Performance

• Discrimination: The GBSA model achieved an Antolini's C-index of 0.680 ± 0.006 and a 1-year cumulative/dynamic AUC of 0.824 ± 0.006.
• Calibration: Predicted survival curves closely tracked Kaplan-Meier (KM) estimates (Mean Absolute Error = 0.010).
  • Risk Stratification: Stratifying patients into low/mid/high risk tertiles based on 1-year PFS probability showed clear prognostic separation.
  • Subtype Robustness: Calibration held across different subtypes (HR+/HER2-, HR-/HER2+, TNBC) and line numbers, despite significant cohort imbalances.
• Time-Dependence: Performance (cdAUC) improved with longer horizons (up to 2 years), while the C-index declined slightly over time, consistent with the diminishing prognostic value of baseline features.

B. Interpretability (SHAP Analysis)

• Dominant Drivers: The model relies primarily on biological signals:
  • Recent metastatic burden (radiology evidence of tumor presence).
  • Tumor marker kinetics (CA15–3, CEA).
  • Genomic mutations (e.g., TP53).
  • Administration of less common chemotherapy agents.
• Non-Drivers: The model showed limited dependence on surveillance cadence or subtype labels, confirming it learned disease dynamics rather than data collection artifacts.
• Risk Stratification: High-risk patients were characterized by liver involvement, high tumor markers, and ECOG status; low-risk patients were characterized by eligibility for endocrine/targeted therapy and absence of recent burden.

C. Robustness Checks

• Ablation: Removing any single modality caused only modest performance degradation. Imaging and planned regimen features provided the strongest standalone signals.
• Leakage Analysis: Models trained without masking late-start agents showed slightly better performance but failed to represent true baseline prognosis (introducing immortal time bias). The masked model was chosen for clinical validity.

4. Key Contributions

1. Auditable EHR Pipeline: A transparent, heuristic-driven workflow to reconstruct mLoTs and label PFS from unstructured radiology reports, addressing the "sparse evidence" problem in later-line mBC.
2. Generalizable Risk Stratification: A single model trained on a heterogeneous cohort that produces well-calibrated risk estimates across diverse subtypes and treatment lines, outperforming subtype-specific models in terms of generalizability.
3. Decision-Point Validity: A rigorous feature engineering strategy that explicitly prevents information leakage (via treatment masking and surveillance suppression), ensuring predictions are actionable at the moment a new therapy line is planned.
4. Clinical Interpretability: Demonstration that the model's risk drivers align with clinical intuition (metastatic burden > subtype labels), supporting its use in clinical decision support.

5. Significance and Limitations

Significance:

• Clinical Utility: Enables proactive planning for subsequent lines of therapy and surveillance intensity based on individualized risk.
• Research Impact: Provides a scalable method to generate real-world evidence for later-line therapies where randomized trials are rare.
• Methodological Advance: Demonstrates that multimodal ML can generalize across heterogeneous real-world data if leakage is strictly controlled.

Limitations:

• Data Granularity: EHR data lacks dose intensity details and specific rationales for drug switches.
• First-Line Bias: The model systematically underestimates PFS for first-line metastatic therapy (mLoT1), likely due to missing pre-metastatic history features.
• Static Snapshot: The model does not incorporate within-line dynamics (e.g., response to treatment, toxicity) after the line starts, limiting its ability to predict outcomes for patients who switch regimens mid-line.

Future Directions:
The authors suggest integrating dynamic time-series models to update risk within a line and utilizing weak supervision to improve mLoT segmentation boundaries.