📊 statistics

biniLasso: Automated cut-point detection via sparse cumulative binarization

The paper introduces biniLasso and sparse biniLasso, novel methods for high-dimensional survival analysis that leverage cumulative binarization and $L_1$ -penalized Cox models to efficiently detect multiple prognostic cut-points per feature with improved interpretability, speed, and prediction accuracy compared to existing techniques.

Original authors: Abdollah Safari, Hamed Halisaz, Peter Loewen

Published 2026-02-13

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Abdollah Safari, Hamed Halisaz, Peter Loewen

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 a doctor trying to figure out which patients are at high risk of a disease based on a continuous number, like their blood pressure or a specific gene level.

The problem is that numbers are messy. Is a blood pressure of 120 bad? What about 121? 130? In the real world, doctors need simple rules to make decisions, like "If blood pressure is above 130, start medication." But finding that perfect "130" number is hard, especially when you have thousands of different numbers (genes, age, weight) to look at at the same time.

This paper introduces two new tools, biniLasso and miniLasso, to solve this puzzle. Here is how they work, explained simply:

1. The Old Way: The "One-Hot" Confusion

Imagine you have a long ruler representing a patient's gene level. Old methods tried to chop this ruler into many tiny, separate slices (like cutting a loaf of bread into individual slices) and then asked, "Which slice is dangerous?"

The problem: This creates a huge mess of data. It's like trying to find a needle in a haystack where the haystack is made of other needles. It's slow, computationally heavy, and often gets confused by the noise.

2. The New Way: The "Cumulative" Ladder

The authors' new method, biniLasso, changes the game. Instead of chopping the ruler into separate slices, imagine a ladder.

You place rungs on the ladder at different heights.
The first rung asks: "Is the value above this low point?"
The second rung asks: "Is the value above this medium point?"
The third rung asks: "Is the value above this high point?"

This is called Cumulative Binarization. It's like stacking buckets. If a patient's value is high enough to reach the top bucket, they automatically fill all the buckets below it. This structure is much easier for the computer to understand and allows it to find the "danger zones" (cut-points) much faster and more accurately than the old "bread-slicing" method.

3. The Two Tools: The Detective vs. The Editor

The paper offers two versions of this tool:

biniLasso (The Detective): This version is great for exploration. It looks at the data and says, "Hey, there might be a danger zone here, and another one there, and maybe a third one!" It finds multiple cut-points to map out complex, winding relationships. It's like a detective who finds every clue, no matter how small, to understand the full story.
miniLasso (The Editor): Sometimes, too many cut-points are confusing for a doctor. "If you have to remember three different thresholds, you'll forget them in a clinic." miniLasso is a "sparse" version. It takes the detective's findings and edits them down. It uses a special trick (called uniLasso) to keep only the most important cut-points. It says, "We can ignore the tiny details; let's just keep the two biggest danger zones." This gives you a simpler, cleaner model that is easier to use in real life without losing much accuracy.

4. Why is this a Big Deal?

Speed: The old methods were like driving a tractor through a field. These new tools are like a sports car. They are 2 to 8 times faster.
Accuracy: In tests using real cancer data (from thousands of patients), these tools found the "danger zones" better than previous methods. They were even able to match the performance of complex, "black box" machine learning models but with the added benefit of being easy to explain.
Real-World Use: The authors tested this on three major types of cancer (Breast, Brain, and Kidney). They found that by using these tools, doctors could turn complex gene numbers into simple rules (e.g., "If Gene X is above 50, risk is high") that predict patient outcomes just as well as complex math models.

The Bottom Line

Think of biniLasso and miniLasso as a smart translator. They take the confusing, continuous language of biology (where everything is a smooth curve) and translate it into the simple, binary language of clinical decisions (Yes/No, High/Low).

biniLasso finds all the hidden thresholds.
miniLasso picks the best ones so doctors can actually use them.

This helps move medical research from "we have a complex model" to "here is a simple rule that saves lives."

1. Problem Statement

In clinical and epidemiological research, discretizing continuous predictors (e.g., converting blood pressure into "high/low" categories) is essential for creating interpretable models. However, existing methods for detecting optimal cut-points in high-dimensional data face significant limitations:

Computational Inefficiency: Methods based on multiple testing are slow and often restricted to a small set of candidate thresholds.
Scalability Issues: Many approaches handle only single predictors or single cut-points per feature, failing to address the joint selection and thresholding required for high-dimensional datasets (e.g., genomics).
Model Limitations: The current state-of-the-art method, binacox, uses one-hot encoding with total-variation (binarsity) penalties. This approach assumes abrupt risk changes (step functions), struggles with gradual biological effects, requires explicit linear constraints for identifiability, and is sensitive to uneven covariate distributions.

2. Methodology

The authors propose biniLasso and its sparse variant, miniLasso, which integrate cumulative binarization with L1 penalization (Lasso) within the Cox proportional hazards framework (and extendable to Generalized Linear Models).

A. Cumulative Binarization

Unlike standard one-hot encoding (which creates disjoint bins), biniLasso employs a cumulative binarization scheme:

Mechanism: For a continuous feature $X_j$ , the method generates a series of binary dummy variables corresponding to splitting points (cut-points). A variable $X_{CB, j, l}$ equals 1 if $X_j > \mu_{j,l}$ and 0 otherwise.
Structure: This creates a nested structure where intervals for larger cut-points are contained within those for smaller ones.
Advantage: This allows for the direct estimation of effect sizes for "lower vs. all higher" comparisons. The resulting design matrix is structurally full-rank (unlike binacox), eliminating the need for auxiliary linear constraints and ensuring a unique, sparse solution via standard L1 regularization.

B. The biniLasso Algorithm

Objective: Minimize the negative log-partial likelihood of the Cox model with an L1 penalty on the coefficients of the binarized features.
Optimization: The problem is formulated as a strictly convex optimization, solvable efficiently without the complex constraints required by binacox.
Flexibility: It can detect multiple cut-points per feature, capturing complex, non-linear relationships.

C. The miniLasso Variant (Sparse & Interpretable)

To address the issue of correlated predictors inherent in cumulative binarization and to enforce parsimony, miniLasso incorporates a two-stage uniLasso regularization:

Stage 1: Fit univariate Cox models for each binarized indicator to generate linear predictor functions.
Stage 2: Compute Leave-One-Out (LOO) predictions to reduce overfitting.
Stage 3: Fit a non-negative Lasso Cox model using the LOO predictions as features.

Benefits: This ensures sign consistency (multivariable coefficients preserve the direction of univariate effects) and produces a more parsimonious model with fewer cut-points while maintaining predictive accuracy.

D. Controlling Cut-Point Count

To meet clinical requirements for a limited number of categories, the authors propose a two-step procedure:

Screening: Independently fit Lasso models for each predictor to rank potential cut-points by influence.
Selection: Retain only the top $m$ cut-points per predictor and fit a final global Lasso (or Group-Lasso) model. This ensures a pre-specified maximum number of cut-points without the computational instability of iterative weight tuning.

3. Key Contributions

Novel Framework: Introduction of cumulative binarization, which shifts the paradigm from disjoint bins to nested intervals, better capturing gradual biological effects and improving identifiability.
Computational Efficiency: The methods are 2–8 times faster than binacox due to the avoidance of constrained optimization and the use of efficient L1 solvers.
Dual Variants:
- biniLasso: Optimized for exploratory analysis to uncover complex, multi-cut-point relationships.
- miniLasso: Optimized for clinical utility, delivering simpler, more interpretable models with comparable accuracy.
Software: Development of the biniLasso R package for reproducibility.

4. Results

Simulation Studies

Speed: Both biniLasso and miniLasso significantly outperformed binacox in computation time across various sample sizes ( $n=300$ to $4000$) and dimensions ( $P=2$ to $100$).
Accuracy:
- In scenarios with sharp thresholds, biniLasso showed lower bias in cut-point estimation compared to binacox.
- In scenarios with gradual "cut-regions" (smooth hazard changes), biniLasso achieved prediction accuracy (Integrated Brier Score) comparable to the true continuous model, whereas binacox performed significantly worse.
- When constrained to exactly two cut-points (Scenario 4), miniLasso and biniLasso maintained high predictive accuracy, while binacox struggled to guarantee the exact number of cut-points.

Real-World Applications (TCGA Datasets)

The methods were applied to three high-dimensional cancer genomics datasets: Breast Invasive Carcinoma (BRCA), Glioblastoma (GBM), and Kidney Renal Clear Cell Carcinoma (KIRC).

Performance: biniLasso and miniLasso consistently outperformed binacox in terms of AIC, IBS, and Concordance Index (C-index). In some cases, the discretized models matched or slightly exceeded the performance of a flexible continuous Generalized Additive Model (CGAM) benchmark.
Cut-Point Characteristics:
- binacox tended to place cut-points at inflection points where hazard trends changed direction.
- biniLasso/miniLasso placed cut-points in regions where the hazard function slope was steepest (where predictions change most rapidly), which is often more clinically relevant.
Parsimony: miniLasso successfully identified fewer cut-points than biniLasso while maintaining similar predictive performance, making it ideal for creating simplified clinical risk categories.

5. Significance and Conclusion

This paper addresses a critical gap in high-dimensional survival analysis by providing a computationally efficient, statistically robust method for automated cut-point detection.

Clinical Utility: The ability to constrain the number of cut-points allows researchers to translate complex genomic data into actionable, simple clinical rules (e.g., "Risk is high if Gene X > Threshold A and Gene Y < Threshold B").
Interpretability vs. Accuracy: The study demonstrates that one does not need to sacrifice predictive accuracy for interpretability; miniLasso achieves a balance where simplified models perform nearly as well as complex continuous models.
Generalizability: While demonstrated on Cox models, the framework is extensible to other Generalized Linear Models (GLMs), broadening its applicability to binary, count, and continuous outcomes.

The authors conclude that a hybrid workflow—using smooth models (like CGAM) to identify relationships and then applying biniLasso/miniLasso to discretize them—is the optimal strategy for balancing statistical power with clinical interpretability.