The clinical anchor of the study is the RigiScan device, the most widely used standardized instrument for NPTR monitoring. Patients wear the device overnight while it records penile rigidity and tumescence during sleep. According to interpretation criteria adopted by national and international guidelines, an effective nocturnal erection is defined as at least one episode with tip rigidity of at least 60 percent sustained for at least 10 minutes. If a patient meets this criterion during monitoring, organic pathology is considered unlikely; if the criterion is not met, an organic cause becomes more probable. The logic rests on the fact that nocturnal erections occur during natural sleep, driven largely by parasympathetic output and central nervous system regulation, and are therefore less contaminated by the performance anxiety that can distort awake, clinic-based assessments.

Despite this advantage, RigiScan testing has practical drawbacks. The equipment is costly, each assessment requires two consecutive nights of recording, and the procedure demands controlled environmental conditions. These constraints make it difficult to deploy as a routine screening tool, which motivated the research team, led by authors from Jiangsu Province Hospital of Chinese Medicine and collaborating institutions, to ask whether simple, routinely collected clinical indicators could estimate the probability that a patient would fail the NPTR criterion. If such a model were reliable enough, it could help stratify patients in settings where RigiScan is not readily available, reserving overnight monitoring for those at highest risk of an abnormal result.

Between June 2024 and December 2024, the investigators consecutively screened 545 men with erectile dysfunction at the inpatient department of Jiangsu Province Hospital of Chinese Medicine. Twenty-six were excluded for reasons including neurogenic erectile dysfunction, a history of pelvic surgery, severe mental illness, or recent use of hormonal or antiandrogen medications. The final cohort of 519 patients, all aged 18 to 70 with a stable partner and a chief complaint of erectile difficulty, completed two consecutive nights of RigiScan monitoring. On each monitoring night, patients received a standardized traditional Chinese medicine foot bath at approximately 21:00, were instructed to sleep at around 22:00, and wore the device for roughly eight hours, of which six hours counted as valid recording time. Experienced technicians who were blinded to all clinical data reviewed the raw waveforms to exclude artifacts from body position changes or device loosening.

The cohort split into 389 men who achieved an effective nocturnal erection and 130 who did not. At baseline, the non-effective group was older, had a higher BMI, and contained larger proportions of smokers and patients with hypertension. Their laboratory panels also showed significantly higher fasting glucose, HbA1c, neutrophil count, and follicle-stimulating hormone. Notably, the two groups did not differ significantly in alcohol consumption, blood lipids, uric acid, liver or kidney function markers, total testosterone, estradiol, or prolactin. As expected, the non-effective group showed markedly fewer erectile events, lower tip and base rigidity, reduced tumescence, and shorter effective erection duration on RigiScan, and penile Doppler ultrasound indicated poorer arterial supply function in that group as well.

To build the predictive models, the researchers divided the data 7:3 into a training set of 364 patients and an internal validation set of 155, performing all feature selection exclusively within the training set to prevent information leakage. Three selection methods were applied and their intersection taken: univariate logistic regression, the Boruta algorithm based on random forests, and least absolute shrinkage and selection operator (LASSO) regression. This conservative three-way agreement narrowed a longer candidate list down to six predictors: HbA1c, fasting glucose, smoking, hypertension, age, and BMI. Crucially, RigiScan parameters themselves and Doppler ultrasound measurements were excluded from the candidate predictors, since the goal was to predict the monitoring outcome from variables available without the overnight test.

Five classification models were then trained with fixed default parameters: logistic regression, random forest, support vector machine, LightGBM, and XGBoost. In the internal validation set, logistic regression achieved the highest area under the receiver operating characteristic curve (AUC) at 0.779 (95 percent confidence interval 0.692 to 0.861), followed by LightGBM at 0.718, XGBoost at 0.700, random forest at 0.699, and the support vector machine at 0.646. DeLong’s test showed that logistic regression significantly outperformed random forest (p = 0.028) and the support vector machine (p = 0.009), though the differences from the two gradient boosting methods were not statistically significant. Logistic regression also posted the lowest Brier score (0.151) and a calibration slope of 1.258, closest to the ideal value of 1, indicating that its predicted probabilities were the best aligned with observed outcomes. In a demanding 5-by-3 nested cross-validation, in which feature selection and model training were repeated inside each training fold, logistic regression again led with a mean AUC of 0.726 plus or minus 0.052 versus 0.666 plus or minus 0.034 for random forest.

To open the black box, the team applied SHAP (SHapley Additive exPlanations) analysis to the winning logistic regression model. BMI, fasting glucose, and age ranked as the top three contributors, followed by hypertension, HbA1c, and smoking. The dependence plots revealed approximately monotonic positive associations between BMI, glucose, and HbA1c and the predicted probability of failing the NPTR criterion, with no clear threshold effect. The authors cautioned that an inflection point around an HbA1c of 6.0 percent reported previously in tree-based models may reflect model-specific behavior rather than a genuine biological threshold. Still, the pattern suggests that even HbA1c values below the diagnostic threshold for diabetes may be associated with the microvascular and endothelial dysfunction that underlies erectile impairment.

The mechanistic backdrop is well described in the paper. Erectile function depends on the coordinated interplay of neurovascular and psychological factors, with the final common pathway being the contraction-relaxation balance of cavernous smooth muscle. Persistent hyperglycemia has been linked to impaired vasodilatory signaling, excessive smooth muscle contraction, and veno-occlusive dysfunction, operating through pathways including the nitric oxide/cyclic guanosine monophosphate cascade, the RhoA/ROCK pathway, oxidative stress, and programmed cell death. At the molecular level, hyperglycemia has been associated with increased formation of advanced glycation end products, suppressed endothelial nitric oxide synthase phosphorylation, reduced nitric oxide bioavailability, enhanced vasoconstrictive effects of endothelin-1 and angiotensin II, and increased release of inflammatory cytokines such as tumor necrosis factor-alpha. Obesity, meanwhile, has been implicated in endothelial dysfunction through adipokine dysregulation, chronic low-grade inflammation, and an unfavorable shift in the testosterone-to-estrogen ratio.

An intriguing contrast emerged when the team compared their model with an earlier one built on intracavernosal injection combined with penile Doppler ultrasound. That pharmacologically stimulated test retained age, hypertension, and testosterone as its main predictors, reflecting vascular structural reserve and age-related arteriosclerosis. The RigiScan-based model instead elevated BMI and the glycemic indicators, and testosterone did not make the final six-variable cut. The authors interpret this as evidence that the two examinations capture different pathophysiological dimensions: Doppler after drug-induced vasodilation probes vascular structure, while nocturnal monitoring during natural sleep integrates central neural drive, sleep architecture, and long-term metabolic adaptation of cavernous tissue. Consistent with this, the anxiety score (HAMA) was retained in the Doppler model but excluded from the RigiScan model, possibly because awake, injection-based testing is more susceptible to anxiety-mediated sympathetic activation than sleep-time monitoring.

The authors are careful about what the model can and cannot do. No model exceeded an AUC of 0.80, meaning discrimination was only moderate, and a screening tool built from routine blood tests cannot yet replace objective RigiScan monitoring. Positive predictive value for logistic regression was 0.469 against a negative predictive value of 0.849, so a reassuring prediction carries more weight than an alarming one. The study is also single-center, retrospective, and cross-sectional, so the findings represent associations rather than causal effects, and external validation is lacking. Sleep quality was not systematically assessed, and data on medications affecting sleep or erection were not collected, leaving room for confounding. Excluding patients with neurogenic erectile dysfunction or pelvic surgery may further limit generalizability. Still, the authors argue that the six metabolic and lifestyle indicators can support preliminary risk stratification, and that when predicted risk is high or clinical suspicion of an organic cause is strong, RigiScan monitoring should still be recommended. The modifiable nature of BMI, glucose, and HbA1c also points toward integrating glycemic control and weight management into comprehensive care for men with erectile dysfunction.

Subject of Research: Machine learning analysis of routine clinical indicators associated with failure to meet RigiScan nocturnal penile tumescence and rigidity criteria in men with erectile dysfunction

Article Title: Identification of factors associated with failure to meet nocturnal penile tumescence and rigidity criteria based on RigiScan monitoring and machine learning algorithms: a cross-sectional study

Article References: Wang, J., Shan, Y., Li, S., Sun, Z., Piao, Z., Zhang, X., & Zhang, M. (2026). Identification of factors associated with failure to meet nocturnal penile tumescence and rigidity criteria based on RigiScan monitoring and machine learning algorithms: a cross-sectional study. Frontiers in Endocrinology, 17, Article 1926525. https://doi.org/10.3389/fendo.2026.1926525

Image Credits: AI Generated

DOI: 10.3389/fendo.2026.1926525

Keywords: erectile dysfunction, RigiScan, nocturnal penile tumescence, machine learning, logistic regression, SHAP, HbA1c, BMI, hyperglycemia, hypertension, cross-sectional study, metabolic syndrome