🤖 AI

Can-SAVE: Deploying Low-Cost and Population-Scale Cancer Screening via Survival Analysis Variables and EHR

Can-SAVE is a lightweight, scalable AI system that leverages survival analysis and electronic health records to significantly outperform conventional screening protocols in cancer detection rates and population coverage, as demonstrated by extensive evaluations on over 2.5 million patients in Russia.

Original authors: Petr Philonenko, Vladimir Kokh, Pavel Blinov

Published 2026-04-29

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Petr Philonenko, Vladimir Kokh, Pavel Blinov

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to find a few specific, hidden needles in a massive haystack. In the medical world, these "needles" are early-stage cancers, and the "haystack" is the entire population of a city or country.

Traditionally, doctors try to find these needles by asking everyone to come in for a very expensive, time-consuming, and complex check-up. This is like hiring a team of people to manually sift through every single piece of hay one by one. It's too slow, too costly, and many needles get missed because you simply can't check everyone.

Enter Can-SAVE: The "Smart Metal Detector"

The paper introduces a new system called Can-SAVE. Instead of hiring more people to sift through hay, Can-SAVE is like a super-smart metal detector that scans the haystack and tells you exactly which 100 pieces of hay are most likely to contain a needle.

Here is how it works, broken down into simple concepts:

1. The Secret Ingredient: Your Medical "Receipts"

Most high-tech AI systems need special data, like genetic tests or complex blood work, to work. Can-SAVE is different. It only looks at your Electronic Health Records (EHR).

Think of your medical history as a long receipt book from a grocery store. It doesn't contain fancy secrets; it just lists what you bought (diagnoses) and when you visited the store (medical services). Can-SAVE reads these "receipts" (using standard codes like ICD-10) to spot patterns that humans might miss.

2. The Two-Step Magic Trick

The system uses a clever combination of two methods to make its predictions:

Step A: The "Weather Forecast" (Survival Analysis)
Imagine a weather forecaster who knows that, statistically, rain is more likely in November than in July. Can-SAVE uses a similar logic called Survival Analysis. It looks at the "weather patterns" of millions of people to understand the general risk of cancer at different ages and for different genders. It creates a baseline "risk map" for the whole population.
Step B: The "Detective" (Machine Learning)
Then, it brings in a detective (a Gradient Boosting Machine). This detective looks at your specific "receipts." Did you visit the doctor often last year? Did you have a specific minor illness? The detective combines your personal story with the "weather forecast" from Step A.

By mixing the big picture (population trends) with the personal details (your specific history), the system creates a "risk score" for every single person.

3. The Results: Finding More Needles, Faster

The researchers tested this system in Russia with massive amounts of data:

The Retrospective Test (Looking Back): They looked at 1.9 million people from the past. When they used Can-SAVE to pick the top 1,000 people to check, they found 4 to 10 times more cancers than the standard screening methods did. It was like the metal detector finding 10 needles where the old method only found 1.
The Real-World Pilot (Looking Forward): They ran a year-long experiment with 426,000 real patients.
- Traditional Method: Found 1,123 cancers.
- Can-SAVE Method: Found 2,148 cancers.
- The Win: They found nearly double the number of cancers without needing any extra money, new machines, or more doctors. They just changed who they invited to the doctor first.

4. Why It's a Big Deal

It's Cheap and Fast: The system is so lightweight that it can process a whole city of 1 million people in less than three hours using standard computer servers. It doesn't need supercomputers.
It's Fair: It works with the data that almost every hospital already has. You don't need special genetic tests or expensive biomarkers.
It Saves Lives: By finding cancers earlier (before symptoms appear), it helps treat patients when the disease is easier to manage.

The Bottom Line

Can-SAVE proves that you don't need futuristic, expensive technology to save lives. By using smart math to read the "receipts" of our daily medical visits, we can build a system that spots cancer risks early, efficiently, and on a massive scale. It turns a chaotic haystack into a sorted list, ensuring the right people get checked at the right time.

1. Problem Statement

Cancer screening faces a critical bottleneck: traditional methods are costly, labor-intensive, and difficult to scale at the population level. Although AI has shown promising advances in cancer detection, most existing systems rely on specialized data (genetics, biomarkers, lifestyle surveys) or complex infrastructures (imaging, NLP), making them impractical for widespread deployment in resource-constrained public health systems.

There is a significant gap between theoretical AI advances and clinical usability. Existing solutions often fail to effectively utilize routine Electronic Health Records (EHR) for population-level risk-based stratification without requiring data beyond standard medical service metrics (ICD-10) and visit histories. The challenge is to create a system that:

Is cost-effective: Requires no specialized hardware or biomarkers.
Is scalable: Capable of processing millions of patients.
Is data-agnostic: Functions with standard EHR data available in almost every clinic.

2. Methodology: The Can-SAVE Framework

Can-SAVE is a lightweight AI system designed to rank population-wide cancer risks exclusively based on routine medical event histories (diagnoses and procedure codes). The core innovation lies in the integration of survival analysis results into a gradient boosting framework.

A. Problem Formulation

The task is formulated as a binary classification problem with temporal constraints:

Input: A patient's EHR ( $Q_i$ ) containing chronologically ordered medical events (ICD-10 codes, procedure codes, dates).
Goal: Predict whether a cancer diagnosis (ICD-10 C00-C97) will occur within the next 12 months ( $t_{pred}$ to $t_{pred} + 12M$ ).
Objective: Maximize the ranking of high-risk patients to prioritize limited screening resources.

B. The Three-Stage Pipeline

Feature Engineering:
- Generation of >700 candidate features from raw EHR data.
- Categories include: Sociodemographics (age, gender), temporal patterns (visit frequency, seasonality), clinical utilization (ratio of diagnoses to procedures), and frequencies of medical events (counts per ICD-10 group).
Construction of Survival Features (The Novelty):
- Instead of using raw EHR data directly for survival modeling, the system extracts population-wide and personalized survival statistics to serve as features for the final classifier.
- Population-wide: Uses Kaplan-Meier estimators to calculate general survival probabilities ( $\hat{S}_{KM}$ ) and gender-stratified survival probabilities. Additionally, "risk gradients" are calculated (the change in survival probability between age $t$ and $t+1$ ).
- Personalized: Uses an Accelerated Failure Time (AFT) model to generate patient-specific survival trajectories and risk gradients based on the patient's specific covariate history.
Integration of Gradient Boosting:
- The outputs of the survival models (probabilities and gradients) are concatenated with standard ML features.
- A CatBoost model (Gradient Boosting Machine) is trained on this extended feature set to predict $P(\text{Cancer} | \text{EHR})$ .
- Rationale: Survival outputs act as regularizers by feeding pre-existing population knowledge into the model to improve generalization and handle the bias-variance trade-off.

C. Baselines

The authors compared Can-SAVE with 17 baselines, including:

Classical ML (Logistic Regression, Random Forest, GBM).
Pure survival models (AFT, Random Survival Forest, DeepHit).
Deep Learning (RNNs, Transformers like BERT/Longformer).
LLM-based approaches (Fine-tuned LLMs, Embedding-based models, LLM summaries).

3. Main Contributions

Novel Architecture: It was demonstrated that combining survival analysis variables with gradient boosting significantly outperforms standalone survival models and complex Deep Learning/LLM approaches for cancer risk ranking using only routine EHR data.
Massive Real-World Validation: The system was validated on a dataset of 2.5 million adults from five Russian regions, representing one of the largest real-world implementations of AI-driven cancer risk assessment.
Prospective Clinical Pilot: A 12-month prospective study supervised by oncologists involving 426,000 patients was conducted to move beyond retrospective analysis and measure actual clinical impact.
Scalability & Efficiency: It was shown that the system can process a city-wide population (1 million patients) in under 3 hours on standard hardware, without requiring new data collection or specialized infrastructures.

4. Experimental Results

A. Retrospective Performance (Numerical Experiments)

Dataset: 175,441 patients (2017–2021).
Metric: Average Precision (AP) was prioritized over ROC-AUC due to extreme class imbalance (cancer is rare).
Results:
- Can-SAVE: Achieved AP = 0.228 (highest value among all methods).
- Best Baseline (LoRA-tuned LLM): AP = 0.193.
- Comparison: Can-SAVE outperformed the best LLM baseline by approximately 18% in AP and performed better than all other classical and Deep Learning methods.
- Ablation Study: Confirmed that the fusion of GBM and survival features is critical; standalone GBM (AP 0.160) and standalone AFT (AP 0.117) were significantly weaker.

B. Oncologist-Supervised Retrospective Study

Scope: 1.9 million patients in 5 regions.
Result: When screening the top 1,000 high-risk patients identified by Can-SAVE, the system detected 41–90 cancer cases per 1,000, compared to 9–15 per 1,000 with traditional screening protocols.
Increase: A 4.1- to 10.0-fold increase in detection rate with identical screening volume.

C. Prospective Pilot (Real-World Implementation)

Scope: 426,210 patients over 12 months.
Comparison: Can-SAVE (Test) vs. National Standard Screening (Control).
Main Results:
- Detection Rate: Can-SAVE detected 2,148 cancer cases versus 1,123 in the control group.
- Efficiency: A 91% increase in cancer detection rate with the same clinical capacity.
- Coverage: Can-SAVE identified 75.4% of all cancer cases detected in the region that year, compared to 39.4% for traditional screening (a +36 percentage point expansion in coverage).
- Resource Utilization: No new hardware components or expensive tests were required; the workflow only changed the prioritization of invitations and a short risk questionnaire.

D. Feature Importance

Age was the dominant feature (20.2% importance).
Outputs of the survival model contributed 39.6% to the total importance, validating the core hypothesis that signals derived from survival data are crucial for ranking.
Visit patterns and clinical markers (e.g., immunology services, benign neoplasms) also contributed significantly.

5. Significance and Impact

Paradigm Shift: Can-SAVE proves that "pragmatic AI" utilizing ubiquitous EHR data can outperform complex, data-hungry models (like LLMs) for population-wide screening.
Accessibility: Since the system relies only on standard ICD-10 codes and visit histories, it is deployable worldwide in resource-poor settings, bypassing the need for genetic tests or specialized imaging.
Clinical Usability: The system was successfully integrated into existing public health workflows and demonstrated that AI can directly improve early detection rates and save lives without disrupting clinical processes.
Scalability: The system's ability to process millions of records in hours on standard servers makes it a viable solution for implementing health policy at the national level.

In summary, Can-SAVE offers a reproducible, cost-effective framework for population-wide cancer prevention, bridging the gap between advanced data science and the real-world constraints of public healthcare.