📄 medicine

Machine Learning to Optimize Busulfan Exposure and Predict Survival in HSCT: Toward Personalized Dosing Strategies

This study demonstrates that machine learning models integrating clinical and pharmacokinetic variables outperform traditional methods in predicting busulfan exposure and survival, thereby enabling more precise, personalized dosing strategies for patients undergoing hematopoietic stem cell transplantation.

Original authors: Dorian Protzenko, Laurent Bourguignon, Clara Boérie, Benjamin Bouchacourt, Raynier Devillier, Joseph Ciccolini

Published 2026-07-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Dorian Protzenko, Laurent Bourguignon, Clara Boérie, Benjamin Bouchacourt, Raynier Devillier, Joseph Ciccolini

Original paper licensed under CC BY 4.0 (https://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 your body is a complex factory, and Busulfan is a powerful, specialized cleaning crew sent in to clear out the old machinery (diseased bone marrow) so new, healthy machinery can be installed (a stem cell transplant).

The problem is that this cleaning crew is a double-edged sword. If you send in too few workers, the old machinery stays, and the transplant fails. If you send in too many, they accidentally destroy the factory's walls, causing life-threatening damage. Finding the "Goldilocks" number of workers for each specific factory is incredibly difficult because every factory is built differently.

Currently, doctors use a standard rule of thumb: "Send in 3 workers for every ton of weight." They then watch the factory closely and adjust the number of workers for the next day based on what they see. This is like guessing the right amount of flour for a cake by weight, then tasting it and adding more later.

This paper asks: Can we use a "smart computer brain" (Machine Learning) to get the recipe right from the very first day, rather than guessing and adjusting?

Here is the breakdown of what the researchers found, using simple analogies:

1. The Old Way vs. The New "Smart" Way

The researchers looked at data from 71 adult patients. They built computer models to predict two things:

How the body handles the drug: Will the factory clear the cleaning crew out quickly, or will it get stuck?
The patient's survival: Will the patient survive the next year?

The Result: The old way of guessing (based on simple weight) was like trying to predict the weather by looking at a single cloud. It wasn't very accurate. The new Machine Learning models, which looked at a whole bunch of clues at once (like liver health, kidney function, and blood protein levels), were much better at predicting the outcome. They were like a super-accurate weather forecast that looks at humidity, wind, pressure, and temperature all at once.

2. The "Tolerance" is Personal, Not Universal

The biggest discovery is that there is no single "safe limit" for how much cleaning crew a factory can handle.

The Analogy: Imagine two houses. House A has strong, thick walls (good liver and kidney function). House B has fragile, thin walls (low protein levels in the blood).
The Finding: House A can handle a massive amount of cleaning crew (high drug exposure) without breaking. House B, however, will collapse if the crew gets even slightly too big.
The Data: Patients with low blood protein levels (albumin) started dying if the drug exposure went above a certain low point. But patients with strong protein levels could handle much higher doses safely.

This means a "one-size-fits-all" target dose is actually dangerous. The safe amount depends entirely on the patient's "factory condition" before they even start.

3. The "Peak" Matters as Much as the "Total"

Doctors usually focus on the total amount of drug the body sees over time (like the total hours the cleaning crew worked). But this study found that the peak intensity (how many workers showed up at once) is also a critical clue.

The Analogy: It's not just about how many hours you work; it's about whether you are asked to lift a heavy piano in one go. Even if the total work hours are the same, lifting the piano all at once might break your back, whereas spreading it out is fine.
The Finding: The computer models realized that a high "peak" concentration of the drug was a warning sign for death, independent of the total amount.

4. The "Crystal Ball" for Dosing

The researchers built a tool that acts like a crystal ball. Before giving the first dose, the computer looks at the patient's age, weight, blood tests, and kidney function.

What it does: It predicts exactly how much drug the patient needs to hit the "sweet spot."
The Impact: If the computer had been used in the past, it would have told doctors to lower the dose for patients who were getting too much (and likely to get hurt) and raise the dose for those who were getting too little (and at risk of the transplant failing).
The Catch: The study admits this is a "proof of concept" using past data. It's like testing a new navigation app on a map of a city you've already driven through. It works well on the map, but the researchers say we need to test it on real, live drivers (a new group of patients) before we trust it to drive the car.

Summary

This paper is a pilot study showing that Machine Learning can act as a highly personalized GPS for Busulfan dosing.

Instead of using a generic map (weight-based dosing), this new approach creates a custom route for every single patient based on their specific biological "terrain." It suggests that by looking at the patient's organ health and blood markers before starting, doctors could avoid dangerous over-dosing and ineffective under-dosing, potentially saving more lives. However, the authors are careful to say this is a hypothesis that needs to be proven in a larger, real-world trial before it becomes standard medical practice.

Technical Summary: Machine Learning to Optimize Busulfan Exposure and Predict Survival in HSCT

Problem Statement
Busulfan is a critical conditioning agent for allogeneic hematopoietic cell transplantation (HCT) with a narrow therapeutic window. While therapeutic drug monitoring (TDM) and model-informed precision dosing (MIPD) have improved target attainment, significant gaps remain. Current practices often rely on a fixed mg/kg starting dose, which fails to fully personalize initial therapy. Furthermore, intra-patient variability driven by clearance saturation complicates dose adaptation, and it remains unclear whether a single cumulative exposure target (Area Under the Curve, AUC) is appropriate for all adults or if tolerable exposure depends on baseline organ function. Additionally, cumulative AUC alone has proven to be a weak predictor of survival, suggesting that other metrics, such as peak concentration ( $C_{max}$ ), and patient-specific factors may be critical.

Methodology
This exploratory, single-center retrospective study analyzed data from 71 adult patients who received high-dose intravenous busulfan conditioning between 2015 and 2022. The study employed a multi-stage machine learning (ML) approach using the H2O framework to address three specific objectives:

Survival Prediction: The authors developed models to predict one-year survival. They compared univariate logistic regression, multivariable logistic regression (with stepwise selection, LASSO, and Ridge), and tree-based algorithms (Random Forest and Gradient Boosting Machines). Performance was evaluated using AUC-ROC, AUC-PR, sensitivity, specificity, and SHAP (SHapley Additive exPlanations) values for interpretability.
Exposure Prediction: Two models were constructed to predict busulfan pharmacokinetics prior to the first dose:
- A Clearance Prediction Model using baseline variables (age, weight, BSA, ferritin, prior exposure).
- A Direct AUC Prediction Model using baseline variables and the administered dose.
  These were compared against standard mg/kg dosing followed by Bayesian adaptation.
Personalized Dosing: A final model was developed to predict the optimal starting dose (mg/kg) required to achieve a specific target AUC, trained on observed AUCs and administered doses.

Key input variables included baseline albumin, glomerular filtration rate (GFR), bilirubin, $\gamma$ -GT, ferritin, body surface area (BSA), and weight.

Key Results

Survival Prediction:
- Univariate analysis of cumulative AUC yielded poor predictive power (AUC-ROC: 0.56).
- Multivariable logistic regression improved performance (AUC-ROC: 0.74), identifying albumin, GFR, and bilirubin as significant predictors.
- Tree-based models (Random Forest and GBM) achieved the highest performance (AUC-ROC: 0.80; AUC-PR: 0.75).
- Key Findings: SHAP analysis revealed a U-shaped relationship between cumulative AUC and survival, indicating that both under- and over-exposure increase mortality risk. Crucially, the models identified that the tolerable exposure threshold is patient-specific:
  - Patients with albumin $\ge$ 40 g/L tolerated cumulative exposures up to ~90.5 mg·h/L (22,000 µM·min).
  - Patients with albumin < 40 g/L showed a sharp increase in mortality beyond ~66 mg·h/L (16,000 µM·min).
- Peak concentration ( $C_{max}$ ) was identified as a complementary predictor to cumulative AUC, capturing acute overexposure risk independently.
Exposure and Clearance Prediction:
- The Random Forest clearance model achieved a Root Mean Square Error (RMSE) of 2.2 L/h.
- The Direct AUC prediction model outperformed the standard clinical strategy (mg/kg + Bayesian adjustment). The ML model achieved an RMSE of 850 µM·min for the first dose compared to 1,200 µM·min for the standard method ( $p < 0.05$ ).
- SHAP analysis confirmed that higher weight, BSA, and age increased clearance, while higher ferritin and prior exposure reduced it.
Personalized Dosing Simulation:
- The ML-guided dosing model demonstrated high accuracy in retrospective simulations, with a mean absolute error of 0.13 mg/kg compared to reference population PK doses.
- For patients who were retrospectively overexposed (>120% of target), the model would have recommended a median dose 15.2% lower.
- For underexposed patients (<80% of target), the model would have recommended a dose 10.2% higher.

Significance and Claims
The authors claim that integrating machine learning with existing pharmacokinetic (PK) approaches can refine the initial dosing of high-dose busulfan and improve the prediction of clinical outcomes. The study demonstrates that:

Personalized Targets: A single, universal cumulative AUC target is likely suboptimal; safe exposure levels depend on baseline organ function (specifically albumin and GFR).
Superior Initial Dosing: ML models utilizing baseline clinical variables can predict busulfan exposure and clearance more accurately than standard mg/kg dosing, potentially reducing the risk of early over- or under-exposure before TDM data is available.
Complementary Role: These models are proposed not as a replacement for established MIPD and Bayesian adaptation, but as a complementary layer to support clinicians in identifying high-risk patients and refining starting doses.

Limitations and Future Directions
The paper explicitly notes that these findings are hypothesis-generating due to the retrospective, single-center nature of the study (n=71), which carries a risk of overfitting. The lack of an independent external validation cohort and the heterogeneity of target AUCs across different conditioning protocols limit generalizability. The authors conclude that prospective, multicentre validation is required before these models can be implemented in routine clinical practice.