A prospective multicenter study on plasma metabolomics and machine learning for diagnosing premature ejaculation and
Jiaochen Luan1,2, Wenchuan Shao1,2, Chunlu Xu3
1Department of Urology, Jiangsu Provincial People's Hospital, The First Affiliated Hospital of Nanjing Medical University, Nanjing, Jiangsu 210029, China.
Background:
Premature ejaculation is a prevalent sexual dysfunction, yet the variable patient response to first-line dapoxetine treatment poses a major clinical challenge, highlighting the unmet need for biomarkers to guide diagnosis and therapy.
Aim:
This study aimed to investigate the distinct plasma metabolic profile of primary premature ejaculation (PPE) patients, and to develop machine learning-based diagnostic and therapeutic response prediction models.
Methods:
A multicenter cohort comprising 69 patients with PPE and 51 healthy control (HC) subjects was enrolled. Plasma samples underwent untargeted metabolomic analysis. Differentially expressed metabolites were identified, and pathway enrichment analyses were conducted using Small Molecule Pathway Database and Kyoto Encyclopedia of Genes and Genomes. Three machine learning algorithms-Support Vector Machine, Random Forest, and Least Absolute Shrinkage and Selection Operator regression-were employed to screen biomarkers. Subsequently, targeted metabolomics analysis was used to quantify neurotransmitter levels.
Outcomes:
The primary outcomes included the Premature Ejaculation Diagnostic Tool, the intravaginal ejaculation latency time, and the Clinical Global Impression of Change scale score after a 4-week observation period of on-demand dapoxetine treatment.
Results:
Multivariate analysis revealed clear separations in metabolic profiles between the PPE and HC groups, and between dapoxetine treatment (DT)-Response and DT-No response groups. Pathway analysis indicated significant enrichment in amino acid metabolism pathways for PPE-related differentially expressed metabolites (DEMs). Additionally, DT-Response-related DEMs were associated with D-Amino acid metabolism and Arginine biosynthesis. Machine learning identified a panel of 4 consensus metabolites for diagnosing PPE, achieving an area under the curve (AUC) of 0.995 in the train cohort and 0.917 in the test cohort. For predicting DT response, three metabolites were selected, forming a model with an AUC of 0.905 (train) and 0.811 (test). It is important to note that these promising initial results require further validation in larger, independent cohorts to confirm their generalizability. Furthermore, targeted metabolomics analysis confirmed significant dysregulation of multiple neurotransmitters in the PPE group.
Clinical Implications:
The machine learning-based models we established show robust performance in diagnostic and dapoxetine treatment response prediction.
Strengths & Limitations:
The establishment of the machine learning-based diagnostic and predictive models represents a key strength, though their clinical translation requires further validation in larger cohorts.
Conclusion:
This study delineates distinct metabolic profiles in PPE, establishes robust machine learning-based models for diagnosis and DT-Response prediction, and reveals the involvement of neurotransmitter dysregulation in its pathophysiology.
