📄 medicine

Sexual Dysfunction Risk Assessment in Peritoneal Dialysis: A Machine Learning Study

This study reveals a high prevalence of sexual dysfunction (77.5%) among peritoneal dialysis patients, identifies depression and residual renal function as key predictors, and demonstrates that machine learning models, particularly CatBoost, can effectively assess individual risk to guide clinical management.

Original authors: Li Deng, Dan Zhao, Qiong Cheng, Dashuang Liu, Liang Liu, Jicong Luo

Published 2026-08-05

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Li Deng, Dan Zhao, Qiong Cheng, Dashuang Liu, Liang Liu, Jicong Luo

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 as a high-tech city, where the kidneys act as the master sanitation department, constantly filtering out trash and keeping the water supply clean. When this department goes on strike, the city gets clogged with waste, leading to a condition called chronic kidney disease. To keep the city running, doctors can install a "backup filtration system" called dialysis. One popular version, peritoneal dialysis, uses the body's own lining (the peritoneum) as a natural filter, allowing patients to clean their blood right at home. But just like any major construction project, fixing the plumbing doesn't always fix everything else in the city. Sometimes, the stress of living with a chronic condition, the chemicals building up in the blood, or the medications needed to keep the heart happy can cause other systems to glitch. One of these often-overlooked glitches is sexual dysfunction—a problem where the body's ability to get or stay aroused, or feel pleasure, gets stuck. While it might feel embarrassing to talk about, it's a common issue that affects how people feel about their lives, their relationships, and their overall happiness.

Now, picture a team of detectives from Xinqiao Hospital in China trying to solve a mystery: "Why are so many people on peritoneal dialysis struggling with this specific problem, and can we predict who is at risk?" They didn't just look at the usual suspects like age or diet; they brought in a new kind of detective: Machine Learning. Think of machine learning as a super-smart robot that can look at thousands of clues at once—like blood test results, mood scores, and medication lists—to find patterns that human eyes might miss. In this study, the researchers gathered data from 209 patients (100 men and 109 women) and asked them to fill out detailed questionnaires about their sexual health. They also checked their medical records for everything from how much urine they still produce to how depressed they felt.

The investigation revealed a startling fact: sexual dysfunction is incredibly common in this group. Out of the 209 patients, a whopping 77.5% (that's 162 people) were experiencing some form of sexual difficulty. The rate was almost identical for both men (78.0%) and women (77.1%), proving that this isn't just a "guy problem" but a universal challenge for anyone on dialysis. When the team tried to figure out why, they used both traditional math and their fancy machine learning robots. The traditional math pointed to a few key culprits: how much residual kidney function a patient still had, whether they were married, and the medications they took. But the machine learning models, specifically one called CatBoost, gave the clearest picture of the big picture.

The most important discovery? The biggest predictor of sexual dysfunction wasn't a blood test or a pill; it was depression. The study found that the severity of a patient's depression (measured by a score called SDS) was the single strongest factor driving the risk. It was like the depression was a heavy fog that made everything else look worse. The second most important factor was residual renal function—essentially, how much of their own kidney work the patient was still doing. The more they could still filter on their own, the better their sexual health tended to be. Interestingly, the machine learning also spotted that taking certain heart medications called beta-blockers doubled the risk of sexual problems.

The researchers also noticed that the "why" changed depending on age. For younger patients (40 and under), the problem seemed to be mostly about feelings, money, and relationships—psychosocial factors. But for older patients (over 40), it was more about the physical body: how well their kidneys were working and how much waste was building up in their blood. The machine learning models were good at predicting who was at risk, with the best model getting about 81% accuracy on new, unseen data. However, the authors are careful to say this is a starting point, not a final answer. Because the study only looked at a snapshot in time (a cross-section) and only at one hospital, they can't say for sure that fixing the depression will fix the sexual problem, though it seems like the most promising place to start. They suggest that in the future, these smart computer models could help doctors spot patients who need extra support before the problem gets too deep, treating the whole person, not just the kidneys.

Technical Summary: Sexual Dysfunction Risk Assessment in Peritoneal Dialysis via Machine Learning

Problem Statement
Sexual dysfunction (SD) is a prevalent yet underrecognized complication in patients undergoing peritoneal dialysis (PD), affecting quality of life significantly. While traditional statistical methods have identified various risk factors, they often struggle to capture complex, non-linear interactions within high-dimensional clinical data. Furthermore, SD remains underdiagnosed due to patient embarrassment and clinician discomfort, necessitating more robust, interpretable tools for risk stratification. This study addresses the gap in predicting SD risk in PD patients by integrating machine learning (ML) with explainable artificial intelligence (XAI) to move beyond group-level averages toward individualized risk assessment.

Methodology
The study employed a retrospective cross-sectional design involving 209 PD patients (100 males, 109 females) from a single-center cohort at Xinqiao Hospital (January 2023–April 2025).

Data Collection & Outcomes: Researchers collected demographic, clinical, biochemical, and psychosocial variables. Sexual function was quantified using the International Index of Erectile Function-5 (IIEF-5) for males and the Female Sexual Function Index (FSFI) for females. Three binary outcomes were defined: Male Erectile Dysfunction (ED), Female Sexual Dysfunction (FSD), and Overall SD.
Statistical Analysis: Univariate analysis, Spearman correlation, and multivariate linear/logistic regression were performed to identify independent predictors. Variance Inflation Factors (VIF) were calculated to assess multicollinearity.
Machine Learning Pipeline: Eight ML algorithms (including CatBoost, Random Forest, XGBoost, LightGBM, SVM, and Logistic Regression) were compared. The dataset was split into training (80%) and testing (20%) sets using stratified random sampling. Model performance was evaluated using AUC-ROC, accuracy, sensitivity, specificity, F1-score, and the Youden index, with five-fold stratified cross-validation to assess generalization.
Explainability: SHapley Additive exPlanations (SHAP) were applied to the best-performing model to quantify feature contributions and visualize relationships between predictors and outcomes.

Key Results

Prevalence: The overall prevalence of SD was 77.5% (162/209), with nearly identical rates in males (78.0%) and females (77.1%).
Traditional Regression: Multivariate analysis identified depression score (SDS), residual renal function (RRF), marital status, and $\alpha$ -blocker use as independent predictors for male erectile function. Marital status emerged as the strongest protective factor for overall SD and female FSD in logistic regression.
Machine Learning Performance:
- CatBoost achieved the highest test-set AUC (0.808) for overall SD prediction but exhibited low sensitivity (0.606) and perfect specificity (1.000), suggesting a conservative classification threshold.
- An Ensemble model (Random Forest + XGBoost) offered a balanced performance (AUC = 0.801).
- Random Forest achieved the highest Youden index (0.646), indicating the best trade-off between sensitivity and specificity.
- Note on Generalization: While test-set AUCs were high (0.772–0.808), five-fold cross-validation AUCs were notably lower (0.53–0.62), suggesting the single random split may have yielded optimistic results given the small test set size ( $n \approx 42$ ).
SHAP Interpretability:
- Depression (SDS) was identified as the dominant predictor (mean $|SHAP| = 0.0455$), followed by RRF, $\beta_2$ -microglobulin, and hemoglobin.
- Subgroup Divergence: Younger patients ( $\le 40$ years) were more influenced by psychosocial factors (SDS, income, marital status), whereas older patients ( $>40$ years) showed stronger associations with physiological parameters (RRF, $\beta_2$ -MG).
- Depression Dominance: In patients with significant depression (SDS $\ge 50$ ), physiological factors lost statistical significance, suggesting depression may overshadow biological determinants in this subgroup.

Key Contributions

High Prevalence Confirmation: The study confirms that SD affects approximately three-quarters of PD patients with no significant gender disparity, challenging the notion that SD is primarily a male concern in dialysis populations.
ML Integration in Nephrology: This is the first study to combine multiple ML algorithms with SHAP-based explainability specifically for SD risk prediction in PD patients.
Identifiable Modifiable Factors: The analysis highlights depression as the strongest modifiable risk factor and RRF as a key protective factor, providing actionable targets for clinical intervention.
Individualized Risk Interpretation: By utilizing SHAP, the study demonstrates how to decompose model predictions to the individual patient level, revealing that patients with similar risk probabilities may arrive at that risk through different pathways (e.g., psychological vs. physiological).

Significance and Limitations
The authors claim that while ML models did not dramatically outperform logistic regression in raw predictive metrics, the integration of XAI (SHAP) provides clinical value by clarifying which variables drive individual risk estimates. This approach supports the hypothesis that treating depression may be the most efficient route to improving sexual function in this population.

The paper maintains a modest tone regarding its findings, explicitly acknowledging limitations:

Sample Size: The cohort size (209 patients) is defensible for regression but small for machine learning, particularly given the high feature-to-sample ratio.
Study Design: The cross-sectional nature precludes causal inference.
Generalizability: The single-center setting limits the external validity of the models.
Missing Data: The study did not measure hormonal parameters (testosterone, prolactin, etc.), and missing data required imputation, introducing uncertainty.
Validation: The authors emphasize that external validation is required to determine whether the models capture true signal or noise, noting the discrepancy between test-set and cross-validation performance.

Future work is proposed to involve multicenter prospective studies to validate these predictive models and evaluate the impact of targeted interventions on sexual function outcomes.

Technical Summary: Sexual Dysfunction Risk Assessment in Peritoneal Dialysis via Machine Learning

More like this