📄 medicine

Development and Evaluation of Nomogram- and Decision Tree-Based Risk Prediction Models for Severe Ovarian Hyperstimulation Syndrome in Patients with High Ovarian Reserve

This study demonstrates that a nomogram-based model incorporating seven independent predictors outperforms a decision tree in predicting severe ovarian hyperstimulation syndrome among patients with high ovarian reserve, offering a validated, web-based tool for individualized clinical risk stratification.

Original authors: Lingli Yang, Dejing Wang

Published 2026-08-11

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Lingli Yang, Dejing Wang

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 you are a gardener trying to grow the most beautiful, rare flowers. You have a special, super-fertile soil that makes plants grow incredibly fast and big. But there's a catch: if you water it too much or give it too much fertilizer, the plants might get so swollen and over-enthusiastic that they burst, leak sap everywhere, and even hurt the whole garden. In the world of human reproduction, this "super-fertile soil" is called a high ovarian reserve, and the "bursting" is a condition called Ovarian Hyperstimulation Syndrome (OHSS). It happens when doctors use medicine to help people have babies, and the ovaries react too strongly. For some patients, this reaction can get scary, causing pain, swelling, and even serious health risks.

Doctors have been trying to build a "weather forecast" to predict who might get this syndrome before it happens. Usually, they look at a patient's age or how many eggs they have. But for the "super-fertile" group, the old forecasts aren't always accurate. Scientists have started using two different types of "crystal balls" to predict the storm: one is a Nomogram, which is like a detailed, multi-layered map where you add up points from different clues to get a final score; the other is a Decision Tree, which is like a game of "20 Questions" where you answer yes or no to a series of questions to find your answer. The big question was: which crystal ball is better at predicting the storm for these specific high-risk gardeners?

This paper, written by researchers Lingli Yang and Dejing Wang, sets out to test these two prediction tools. They looked back at the records of 803 patients with high ovarian reserves who were trying to have babies between 2020 and 2025. They split these patients into two groups: a "training" group of 614 people to build the models, and a separate "test" group of 189 people to see if the models worked on new data.

First, the researchers played detective to find the best clues. They started with 20 different pieces of information, like weight, hormone levels, and how many eggs were retrieved. After some serious number-crunching, they narrowed it down to just seven "super-clues" that mattered most:

AMH: A hormone that tells how many eggs are in the bank.
AFC: The actual count of tiny egg sacs seen on an ultrasound.
bLH/bFSH ratio: A specific balance between two hormones that control the ovaries.
Estradiol (E₂) on Day 5: A hormone level measured early in the treatment.
Estradiol (E₂) on Trigger Day: The hormone level right before the final egg retrieval shot.
Follicles ≥ 10 mm: The number of growing egg sacs that reached a certain size.
Oocyte yield: The total number of eggs actually collected.

Using these seven clues, they built both the Nomogram (the point-scoring map) and the Decision Tree (the yes/no question game). Then, they put them to the test.

The results showed that the Nomogram was the clear winner. When they tested it on the training group, it correctly predicted the risk 88.4% of the time (an "AUC" score of 0.884). The Decision Tree was okay, scoring 86.5%, but it wasn't as sharp. When they tested the models on the new group of 189 patients (the "external validation"), the Nomogram stayed strong with a score of 87.4%, while the Decision Tree's score dropped to 79.5%. This suggests the Decision Tree might have been "memorizing" the training data too well and got confused by new patients, whereas the Nomogram remained reliable.

The researchers also checked how well the predictions matched reality. The Nomogram's predictions lined up almost perfectly with what actually happened, while the Decision Tree was a bit off. They even built a web-based calculator (a digital tool anyone can use on a computer) based on the winning Nomogram. Now, a doctor can type in a patient's seven specific numbers, and the calculator instantly gives a percentage chance of the patient developing severe OHSS.

In short, the paper suggests that for patients with high ovarian reserves, the detailed, point-based Nomogram is a more accurate and stable tool for predicting severe OHSS than the simpler Decision Tree. It didn't just find a new way to guess; it proved that one specific method works better than the other for this tricky group of patients, offering a way to spot the danger early and keep the "garden" safe.

Technical Summary: Development and Evaluation of Nomogram- and Decision Tree-Based Risk Prediction Models for Severe OHSS

Problem Statement
Ovarian Hyperstimulation Syndrome (OHSS) remains a serious, potentially life-threatening complication of assisted reproductive technology (ART), particularly for patients with high ovarian reserve who exhibit exaggerated responses to controlled ovarian stimulation (COS). While severe OHSS can lead to multi-organ failure and adverse pregnancy outcomes, current management is largely supportive, and no specific therapy exists. Although machine learning (ML) techniques are increasingly used in disease prediction, existing models for OHSS often rely on conventional logistic regression or lack specific validation for the high-risk subgroup of patients with high ovarian reserve. Furthermore, the comparative utility of different ML approaches, specifically nomograms versus decision trees, within this specific population has not been systematically explored.

Methodology
This retrospective study utilized data from 803 infertile patients with high ovarian reserve (defined as age <35, AFC >20, AMH >3.6 ng/mL) who underwent a long-acting follicle-phase protocol at the Affiliated Hospital of Zunyi Medical University between January 2020 and October 2025.

Data Split: The cohort was divided into a training set (614 patients, 2020–2024) for model development and internal validation, and an external validation set (189 patients, 2025).
Feature Selection: Researchers first conducted univariate analysis to identify 20 variables with statistical significance ( $P < 0.05$ ). These were subsequently processed through Least Absolute Shrinkage and Selection Operator (Lasso) regression with 10-fold cross-validation to isolate independent predictors.
Model Construction: Two distinct models were built:
1. A Nomogram using the rms package in R, translating logistic regression results into a visual scoring system.
2. A Decision Tree using the rpart package, creating a 4-level binary classification tree.
Evaluation Metrics: Model performance was assessed using Receiver Operating Characteristic (ROC) curves (Area Under the Curve - AUC), calibration plots, and Decision Curve Analysis (DCA). Internal validation employed bootstrap resampling (1,000 iterations).
Implementation: A web-based risk calculator was developed using the Shiny package based on the superior performing model.

Key Contributions

Identification of Specific Predictors: The study identified seven independent risk factors specifically for severe OHSS in high-reserve patients: Anti-Müllerian Hormone (AMH), Antral Follicle Count (AFC), basal LH/FSH ratio (bLH/bFSH), Estradiol ( $E_2$ ) on day 5 of Gn stimulation, $E_2$ on the day of hCG trigger, the number of follicles $\ge$ 10 mm on hCG day, and oocyte yield.
Comparative Analysis: The paper provides a direct methodological comparison between a traditional visual regression tool (nomogram) and a classical ML classifier (decision tree) within the same high-risk cohort.
Clinical Tool Development: The study translated the optimal statistical model into a functional, web-based risk calculator to facilitate real-time, individualized risk assessment in clinical practice.

Results

Predictive Performance: In the training cohort, the nomogram demonstrated superior discrimination with an AUC of 0.884 (95% CI: 0.841–0.927) compared to the decision tree's 0.865 (95% CI: 0.814–0.916).
Validation: Following bootstrap internal validation, the nomogram maintained a robust AUC of 0.880, while the decision tree dropped to 0.812. In external validation, the nomogram achieved an AUC of 0.874 (95% CI: 0.821–0.928), whereas the decision tree fell to 0.795 (95% CI: 0.702–0.888).
Calibration and Utility: Calibration curves indicated the nomogram was closer to the ideal line, and Decision Curve Analysis (DCA) showed the nomogram provided a higher net clinical benefit across a wider range of threshold probabilities.
Tool Output: The web-based calculator successfully estimated risk probabilities (e.g., predicting a 90.3% risk for a specific case study) consistent with the nomogram's visual output.

Significance and Claims
The authors claim that the nomogram-based model demonstrates superior predictive performance, reproducibility, and stability compared to the decision tree approach for patients with high ovarian reserve. The study posits that integrating this model with the web-based calculator can facilitate individualized risk stratification and precision management in clinical settings. Specifically, the tool aims to guide clinicians in triaging patients: those with low predicted risk (<10%) may proceed with routine treatment, while those with moderate (10–30%) or high (>30%) risk can receive enhanced monitoring or consider cycle cancellation to prevent severe complications. The authors acknowledge limitations, including the single-center retrospective design and the lack of comparison with other advanced ML algorithms (e.g., XGBoost, Random Forest), suggesting that future prospective, multicenter studies are necessary to further validate these findings.

Technical Summary: Development and Evaluation of Nomogram- and Decision Tree-Based Risk Prediction Models for Severe OHSS

More like this