🤖 AI

Tackling Small Sample Survival Analysis via Transfer Learning: A Study of Colorectal Cancer Prognosis

This study proposes and evaluates transfer learning methods, including a novel Transfer Survival Forest model, to enhance the performance of survival analysis on small-sample colorectal cancer prognosis data by leveraging knowledge from larger source datasets.

Original authors: Yonghao Zhao, Changtao Li, Chi Shu, Qingbin Wu, Hong Li, Chuan Xu, Tianrui Li, Ziqiang Wang, Zhipeng Luo, Yazhou He

Published 2026-06-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yonghao Zhao, Changtao Li, Chi Shu, Qingbin Wu, Hong Li, Chuan Xu, Tianrui Li, Ziqiang Wang, Zhipeng Luo, Yazhou He

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

The Core Problem: Learning with Too Little Data

Imagine you are trying to teach a student how to predict the weather. If you give them 10,000 years of weather records, they will learn very quickly and accurately. But what if you only have weather records for the last two weeks? The student will struggle. They won’t have enough examples to spot the patterns, and their predictions will likely be wrong.

In medicine, this is a common problem. Doctors want to predict how long a cancer patient might survive (this is called "survival analysis"). However, for specific types of cancer or specific hospitals, they often only have a small number of patient records—sometimes fewer than 100. This is called a "small sample" problem. With so little data, computer models usually fail to make accurate predictions.

The Solution: "Cheat Sheets" from Other Sources

The researchers in this paper propose a solution called Transfer Learning.

Think of it like this: If you want to learn to play a new video game, but you’ve never played it before, you might watch a tutorial or play a similar game first. You don’t start from zero; you bring your existing skills with you.

In this study, the researchers used a massive database of cancer patients from the United States (called SEER) as the "tutorial." This database had over 27,000 records. They taught their computer models using this huge dataset. Then, they took those "smart" models and applied them to a much smaller group of patients from a hospital in China (called WCH), which had only 728 records.

The idea is that the model already knows the general "rules" of cancer survival from the big US database, so it needs fewer examples from the Chinese hospital to make good predictions.

The Two Types of "Students" (Models)

The researchers tested this "cheat sheet" method on two different types of computer models:

1. The Neural Networks (The "Black Box" Experts)

These are complex models like DeepSurv, Cox-CC, and DeepHit. Think of these like a master chef who has memorized thousands of recipes.

How Transfer Learning works here: The chef learns basic cooking techniques from a huge international cookbook (the SEER data). When they move to a new kitchen (the WCH data), they can either:
- Retrain: Throw out the old recipes and learn entirely new ones based on the local ingredients.
- Fine-tune: Keep their basic cooking skills but adjust the seasoning to fit the local taste.
The Finding: The researchers found that if the new kitchen has plenty of ingredients (more than 500 patients), it’s better to retrain (learn new recipes). But if ingredients are scarce (fewer than 200 patients), it’s better to fine-tune (keep the old skills and just tweak them).

2. The Random Survival Forests (The "Tree" Experts)

This is a different type of model called Random Survival Forest (RSF). Think of this like a committee of judges making decisions by asking a series of yes/no questions (like a game of "20 Questions").

The Innovation: The researchers created a new method called Transfer Survival Forest (TSF).
How it works: Instead of starting from scratch, the new committee looks at the questions the old committee (from the SEER data) asked. They "transplant" the most useful questions from the old committee to the new one.
The Twist: They don’t just copy the questions; they adjust the answers based on the new patients. For example, if the old committee asked, "Is the tumor larger than 2cm?", the new committee keeps that question but might change the threshold to "Is the tumor larger than 3cm?" based on the new data.

The Results: Did It Work?

Yes, it worked very well. The researchers measured accuracy using a score called C-index (where 1.0 is perfect prediction and 0.5 is random guessing).

Neural Networks: Using the "cheat sheet" method improved the scores. For example, one model (Cox-CC) went from a score of 0.78 (using only local data) to 0.81 (using transfer learning).
Random Survival Forests: This method saw the biggest jump. The score went from 0.79 to 0.83. This was the highest performance of all the models tested.

The "Small Sample" Bonus:
The improvement was even more dramatic when the data was extremely small. When they tested with only 50 patients, the transfer learning methods still performed much better than trying to learn from scratch.

Important Caveats (The "Fine Print")

The paper highlights a few important limitations:

Don't Over-Adjust: For the "Tree" models (TSF), if you have very few patients (less than 40), trying to adjust the questions too much can actually make the predictions worse than if you had just used the original "cheat sheet" without changes. This is called "negative transfer." The researchers suggest keeping the decision trees shallow (simple) when data is scarce.
Same Language Required: This method works best when the data from the source (US) and the target (China) use the same types of information (like age, tumor size, etc.). If the data types are completely different, the "cheat sheet" won't help.

Conclusion

In simple terms, this paper shows that when doctors have too few patient records to make accurate survival predictions, they can borrow knowledge from large, public databases. By pre-training computer models on big data and then carefully adjusting them for local patients, they can get much better results than if they tried to learn from the small local dataset alone. This is especially helpful for rare diseases or small hospitals where gathering large amounts of data is difficult.

Technical Summary: Tackling Small Sample Survival Analysis via Transfer Learning

Problem Statement
Survival analysis is a critical statistical method in medical research for estimating time-to-event outcomes, such as disease progression or mortality. However, the precision and robustness of these analyses are heavily dependent on sample size and event incidence rates. In clinical practice, particularly in oncology, researchers often face "small sample" scenarios due to prolonged follow-up times, low event rates, or the focus on specific molecular subgroups in personalized medicine. Small datasets (often <200 cases) are frequently insufficient to induce useful patterns for survival prediction, leading to overfitting or poor generalization. While large open-access datasets (e.g., SEER, UK Biobank) are increasingly available, leveraging this external knowledge to improve local, small-scale clinical predictions remains a challenge. This study addresses the need for novel analytic approaches that enhance prediction performance in small-sample survival analysis settings.

Methodology
The authors propose transfer learning frameworks for common survival models, categorized into parametric (neural network-based) and non-parametric (tree-based) models. The experimental setting involves a source task using 27,379 Stage I colorectal cancer (CRC) patients from the SEER database and a target task using 728 Stage I CRC patients from West China Hospital (WCH).

Transfer for Neural Network Models (Parametric):
The study applies standard transfer learning techniques to DeepSurv, Cox-CC (Cox-based neural networks), and DeepHit. Two strategies are evaluated:
- Retraining (RT): All network parameters are re-optimized on the target task data.
- Fine-tuning (FT): Only a subset of parameters (specifically the output layer) is re-optimized, while deeper layers representing general feature representations are frozen.
  The authors investigate the relationship between transfer granularity and data availability, hypothesizing that the optimal strategy depends on the size of the target dataset.
Transfer for Random Survival Forests (Non-Parametric):
The primary methodological contribution is a novel Transfer Survival Forest (TSF) model. Unlike neural networks, tree structures cannot be directly fine-tuned via gradient descent. The proposed workflow involves:
- Source Forest Training: A standard Random Survival Forest (RSF) is trained on the source data.
- Structure Extraction: The method identifies frequent tree structures from the source forest. Two transfer principles are defined:
  - TSF-Tk: Counts the frequency of tree structures with the top k levels, capturing feature co-occurrence relationships.
  - DP (Depth-wise Probability): Counts the independent probability of feature occurrence at each depth level, ignoring co-occurrence.
- Target Forest Fine-tuning: Prototype trees are sampled from the source distribution. For the transferred top k levels, the splitting features are preserved from the source, but the splitting values are recalculated using the target data. Lower levels (> k) are grown independently using standard RSF procedures.

Key Contributions

Framework Development: The paper develops and evaluates transfer learning frameworks for both parametric (DeepSurv, Cox-CC, DeepHit) and non-parametric (RSF) survival models.
Novel TSF Model: The authors introduce a new transfer technique for Random Survival Forests that transfers tree structures (feature combinations) from source tasks and fine-tunes them by recalculating splitting values on target data.
Empirical Analysis of Transfer Granularity: The study provides empirical evidence on how the choice between retraining and fine-tuning for neural networks depends on target data size, and how tree depth and transfer strategy affect RSF performance in small-sample regimes.

Results
The models were evaluated using the time-dependent concordance index ( $C_{td}$ ) across various target sample sizes (ranging from 728 down to 50).

Neural Network Models: Transfer learning consistently improved performance over training on target data alone.
- For larger target samples ( $\ge$ 500), retraining generally outperformed fine-tuning.
- For smaller target samples ( $\le$ 200), fine-tuning outperformed retraining, likely because insufficient data leads to suboptimal parameter optimization in full retraining.
- Specific improvements included Cox-CC’s $C_{td}$ increasing from 0.7868 to 0.8111 and DeepSurv’s from 0.7722 to 0.8043 on the full WCH dataset.
Transfer Survival Forest (TSF):
- TSF achieved the highest performance among all tested models. On the full WCH dataset, RSF’s $C_{td}$ increased from 0.7940 (target-only) to 0.8297 (TSF-T3).
- TSF outperformed non-transfer models when training data sizes were $\ge$ 40.
- Negative Transfer: When target data was severely limited (<40 samples), TSF performance dropped below the target-only baseline. This is attributed to the high variance in recalculating splitting values with very few samples. The authors recommend transferring only the top 1–2 levels (TSF-T1/T2) and keeping trees shallow to mitigate overfitting in these extreme cases.

Significance and Claims
The paper claims that properly designed transfer learning techniques can significantly enhance and improve current survival models used for cancer prognosis, particularly in small-sample settings. It demonstrates that:

Source models from large databases (like SEER) provide strong baselines that surpass local small-sample models.
Fine-tuning strategies must be adapted to data availability: retraining is preferable for moderately sized datasets, while fine-tuning is more robust for very small datasets.
Tree-based models, while interpretable, rely more heavily on target data for fine-tuning (splitting values) than neural networks, making them more susceptible to negative transfer in extremely small samples unless constrained (e.g., shallow trees).

The study concludes that transfer learning is a feasible and superior approach for improving cancer prognosis predictions when local data is scarce, bridging the gap between large-scale registry data and local clinical cohorts.