COMBIO: Automated Discovery of a Compact Composite Biomarker from Routine Laboratory Panels

Tomer D. Meirman, Lior S. Rokach, Bracha Shapira, Seffi Cohen, Noa DaganarXiv 2026 · arXiv

TL;DR

COMBIO is an automated pipeline that distills complex machine-learning models into simple, multiplicative formulas for clinical biomarkers. By searching a constrained space of small integer exponents guided by SHAP values, it discovers compact indices that rival the predictive power of standard logistic regression while remaining easy to calculate at the bedside.

What Problem Does This Paper Address?

In clinical settings, there is a persistent tension between performance and practicality. Modern machine-learning models (like gradient-boosted trees) can predict patient outcomes with high accuracy using routine laboratory data. However, these "black box" models are difficult to interpret and even harder to implement without specialized software.

Clinicians often prefer "hand-crafted" indices—like the neutrophil-to-lymphocyte ratio (NLR) or the FIB-4 index for liver fibrosis—because they are transparent, easy to calculate, and provide a single, intuitive number. The challenge is that manually creating these indices is slow and limited to specific clinical questions. COMBIO aims to bridge this gap by automating the discovery of these simple, "index-like" biomarkers from any routine lab panel.

Methods

The COMBIO (Composite Monomial Biomarker) pipeline follows a three-stage process to translate a complex model into a simple formula:

  1. Reference Model Training: A gradient-boosted tree (XGBoost) is trained on the full laboratory panel to learn the underlying predictive signals.
  2. Feature Ranking and Directionality: The researchers use Tree SHAP (Shapley Additive Explanations) to identify the top five most influential laboratory measurements. Crucially, they analyze the SHAP values to determine the "direction" of each variable (e.g., does a higher BUN value increase or decrease the risk?).
  3. Formula Search: The system performs an exhaustive search over a restricted hypothesis space. It looks for a monomial—a product of 2 to 4 variables, each raised to a small integer power between -2 and 2 (e.g., $x^1, x^2, x^{-1}, x^{-2}$).
Figure 1: COMBIO formula discovery, illustrated by MIMIC-IV in-hospital mortality. A model fitted in inner training supplies mean absolute Tree SHAP values and directional trends in inner validation. These guide measurement ranking and eligible integer exponents for formulas with two to four terms. Validation AUROC selects the formula, with risk orientation set using inner training labels. Test data are reserved before fitting and used after selection. BUN, blood urea nitrogen; WBC, white blood cell count; RDW, red cell distribution width; SHAP, Shapley additive explanations. Directional trends describe model associations rather than causal effects. The selected example is C = BUN × glucose × RDW2/albumin2; WBC receives exponent zero.

The final formula is selected based on the highest Area Under the Receiver Operating Characteristic curve (AUROC) on a validation set. Mathematically, this multiplicative form is equivalent to a linear model applied to log-transformed data, a common structure in biological systems.

Datasets

The authors benchmarked COMBIO across six diverse clinical tasks using four major datasets:

  • NHANES (National Health and Nutrition Examination Survey): Used for iron overload screening, 5-year all-cause mortality, and undiagnosed diabetes.
  • ILPD (Indian Liver Patient Dataset): Used for liver disease classification.
  • MIMIC-IV (Medical Information Mart for Intensive Care IV): Used for in-hospital mortality and sepsis case-finding.
  • eICU Collaborative Research Database: Used as an independent external validation set for the two hospital-based tasks (mortality and sepsis).

Key Insights and Ideas

  • The Compactness-Performance Trade-off: The primary goal of COMBIO isn't to beat state-of-the-art ML models in accuracy, but to see how much performance is lost in exchange for extreme simplicity. The authors found the loss is surprisingly minimal.
  • SHAP as a Pruning Tool: Rather than just using SHAP for post-hoc explanation, COMBIO uses it to constrain the search space. By identifying which variables matter and their general trend, the exhaustive search becomes computationally feasible and clinically grounded.
  • Structural Familiarity: By forcing the output into a multiplicative formula (like $C = \text{BUN} \times \text{glucose} \times \text{RDW}^2 / \text{albumin}^2$), the resulting biomarker feels familiar to clinicians who already use ratios and products in daily practice.
  • Transportability: The discovered formulas proved robust. A formula "frozen" in one hospital system (MIMIC) still provided meaningful risk stratification when applied to a completely different system (e.g., eICU), though the specific risk thresholds (calibration) needed adjustment.

Key Results

COMBIO consistently produced biomarkers that performed nearly as well as more flexible models:

  • Competitive Discrimination: Across all six tasks, COMBIO’s AUROC was within 0.03 of a standard five-variable logistic regression.
  • Mortality Prediction: In the MIMIC-IV mortality task, COMBIO achieved an AUROC of 0.773, outperforming several established clinical indices like the BUN/albumin ratio (0.746) and the Prognostic Nutritional Index (0.738).
  • High Coverage: A version of the mortality biomarker restricted to the most widely available labs (platelets, RDW, BUN, etc.) achieved an AUROC of 0.786 in a cohort covering 92% of all hospital admissions.
Figure 2: Test AUROC for COMBIO, five-candidate logistic regression and full XGBoost. Panel A lines are 95% percentile bootstrap intervals for every method; panel B shows the paired COMBIO- minus-logistic AUROC difference intervals. Resampling was by participant or admission, as stated in Methods. AUROC, area under the receiver operating characteristic curve.

Limitations and Future Work

  • Complete-Case Requirement: Because the formulas are products, they require all component lab values to be present. In clinical practice, missing data is common, which could limit the applicability of a specific formula to only a subset of patients.
  • Calibration: While the ranking of patients (discrimination) transfers well across hospitals, the absolute risk scores do not. Local validation and recalibration are essential before these formulas could be used for clinical decision-making.
  • Descriptive, Not Causal: The authors emphasize that these formulas describe statistical associations in the data and should not be interpreted as having a direct causal link to the outcomes.
  • Future Directions: Future work could explore handling missing data more gracefully or expanding the search space to include simple additive terms while maintaining the "pocket-calculator" simplicity of the current approach.

Summary updated

Abstract

Machine-learning models based on routine laboratory panels can predict outcomes but can be difficult to interpret and operationalize. We present COMBIO, an automated pipeline that distils such a model into a compact closed-form biomarker. The composite is a product of laboratory measurements raised to integer powers from minus two to two. The sign of each exponent indicates whether higher values of the corresponding measurement increase or decrease the composite. Shapley additive explanation values from the gradient-boosted reference model rank five candidate measurements and guide eligible exponent signs. An exhaustive search over small integer exponents selects the formula on an inner validation set, before evaluation on the reserved test set. We evaluated COMBIO across six tasks: three National Health and Nutrition Examination Survey outcomes, liver-disease classification in the Indian Liver Patient Dataset, and Medical Information Mart for Intensive Care IV in-hospital mortality and sepsis case-finding, with external evaluation in the eICU Collaborative Research Database for the two hospital tasks. Across the six test sets, COMBIO's area under the receiver operating characteristic curve differed by at most 0.03 from logistic regression using the five candidate laboratory measurements. The hospital mortality composite had an area under the curve of 0.698 in the external database. A version restricted to widely available measurements in the source hospital database achieved an area under the curve of 0.786 in a cohort covering 92% of admissions. The multiplicative representation provides compactness rather than additional accuracy, and local validation and recalibration are required before clinical use.
BibTeX
@misc{meirman2026combio,
  author = {Tomer D. Meirman and Lior S. Rokach and Bracha Shapira and Seffi Cohen and Noa Dagan},
  title = {{COMBIO: Automated Discovery of a Compact Composite Biomarker from Routine Laboratory Panels}},
  year = {2026},
  eprint = {2611.09102},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.09102},
}
Figure