Sexual dysfunction (SD) is one of the most common and distressing consequences of cancer treatment, yet it remains systematically under-addressed in clinical care. Cancer treatments including surgery, chemotherapy, radiation, and hormonal therapy can damage blood vessels, nerves, and hormone systems essential for sexual function.
The scale of the problem is substantial: sexual dysfunction affects up to 80% of men with prostate cancer, 40-80% of women with gynecological cancers, and 55-73% of patients with hematologic cancers. These difficulties in desire, arousal, or orgasm frequently persist throughout survivorship, significantly impacting quality of life, psychological well-being, and intimate relationships.
Despite this prevalence, many patients with cancer receive little to no information or guidance on managing sexual side effects. Time constraints in oncology settings, limited provider training, and discomfort addressing sensitive topics all contribute to this gap in care.
Artificial intelligence (AI) and machine learning (ML) tools are increasingly being explored as a way to bridge this gap, with applications including predicting which patients are at highest risk of sexual dysfunction, extracting unreported symptoms from medical records, and helping plan treatments that better spare sexual function structures.
Researchers conducted a comprehensive, PRISMA-guideline systematic review registered in the PROSPERO international database. They searched PubMed, EMBASE, and Web of Science, screening 3,862 studies using terms including artificial intelligence, machine learning, sexual dysfunction, erectile dysfunction, and sexual health in combination with cancer-related terms.
After thorough screening, 28 studies published from 2002 to 2025 met all inclusion criteria: they had to apply AI or ML models to sexual health prediction or management in cancer populations, involve human subjects, and report model performance metrics. Reviews, conference abstracts, and studies without tested AI models were excluded.
Study quality was assessed using two rigorous frameworks: PROBAST (Prediction model Risk of Bias Assessment Tool) evaluated methodological quality across four domains, and TRIPOD+AI guidelines assessed transparency and completeness of reporting across 27 checklist items specific to AI prediction models.
Data was extracted on study design, cancer type, AI methodology, input variables, outcome measures, and model performance. Because of high heterogeneity in study designs and cancer populations, a formal meta-analysis was not performed. Instead, a narrative synthesis grouped studies by cancer care phase and AI methodology.
Prostate cancer dominated the literature, representing 50% of included studies, followed by cervical cancer (25%) and breast cancer (7%). This reflects both the prevalence of prostate cancer and the availability of large clinical datasets, but also highlights an important research gap: female gynecological cancers are significantly under-studied.
Publications in this space have surged dramatically over time: only 2 studies were published between 2002 and 2009, 6 between 2010 and 2019, and 20 between 2020 and 2025, reflecting the accelerating adoption of AI methods in oncology quality-of-life research.
AI applications spanned five phases of the cancer care continuum. During prevention and screening, AI models predicted cervical and ovarian cancer risk from behavioral and demographic data. During treatment, AI tools supported decision-making and predicted treatment-induced sexual toxicity. During survivorship, AI predicted post-treatment sexual function and quality of life.
Natural language processing (NLP) was applied to extract undocumented sexual health symptoms from electronic health records, revealing outcomes that patients and clinicians rarely document in structured fields. Deep learning was applied to medical imaging to automatically identify and protect anatomical structures critical to sexual function during radiation treatment planning.
Across the 28 studies, the most frequently used algorithm was regression-based models (n=18), followed by gradient boosting machines (GBM) and neural networks (both n=14). Random Forest, support vector machines, decision trees, k-nearest neighbors, and deep learning models were also represented.
Random Forest achieved the highest overall performance with a median AUC of 0.98 (range 0.91-0.99), sensitivity of 0.98, specificity of 0.99, and F1 score of 0.98. Boosting models (GBM, XGBoost, AdaBoost) also performed strongly with a median AUC of 0.94. These ensemble methods consistently outperformed traditional regression approaches.
Logistic regression and traditional regression models showed lower AUC values (median 0.83, range 0.60-0.97). Support vector machines showed the lowest performance with a median AUC of 0.77. However, regression models remain widely used because of their interpretability and clinical familiarity.
For imaging tasks like automatically segmenting neurovascular bundles critical to erectile function, deep learning models achieved Dice similarity coefficients (DSC) of 81%, compared to expert manual segmentation. For the internal pudendal artery, a key vessel for erectile function, deep learning achieved a DSC of 62%, demonstrating the feasibility of AI-guided sparing of these structures during radiation planning.
The most clinically developed AI applications focus on predicting erectile dysfunction (ED) after radical prostatectomy for prostate cancer. A gradient boosting model (GBM) trained on 2,653 patients predicted sexual function at 3, 6, 12, and 24 months after surgery, achieving an AUC of 0.91. This type of dynamic prediction model could help men and their physicians set realistic expectations and plan supportive interventions.
An artificial neural network (ANN) trained on 8,524 patients predicted erectile function 12 months after nerve-sparing robotic radical prostatectomy, achieving an AUC of 0.74. Key input variables included age, BMI, Gleason score, diabetes status, and baseline erectile function scores, showing that preoperative patient characteristics meaningfully predict outcomes.
A logistic regression model with recursive feature elimination in 964 patients predicted ED at 1 and 2 years post-diagnosis with AUCs of 0.84 and 0.81, respectively. These models consistently found that baseline erectile function, nerve-sparing surgical technique, and comorbidities like diabetes were the strongest predictors of post-treatment sexual function.
Beyond prediction, AI is being used to help patients make treatment decisions. A web-based decision-aid tool in 750 prostate cancer patients used patient-specific inputs to rank treatment options by projected sexual and urinary outcomes, demonstrating that no single treatment is universally optimal, and personalized decision support is valuable.
Despite promising AI performance metrics, the overall quality of the evidence is concerning. All 28 included studies (100%) were rated as high risk of bias in the analysis domain by the PROBAST tool. This means every study had methodological weaknesses including overfitting, lack of calibration testing, or absence of external validation that limit confidence in their real-world reliability.
Overall adherence to TRIPOD+AI reporting guidelines was only 60% on average. Critical quality items were almost universally absent: model calibration assessment was present in only 6 of 28 studies (21%), external validation in 4 of 28 (14%), code sharing in 0 of 28 (0%), and protocol registration in only 4 of 28 (14%).
More than half of studies (53.6%) did not use any validated sexual health assessment tool, relying instead on general quality-of-life instruments not designed to capture cancer-specific sexual symptoms. Without standardized outcomes, comparing models across studies or translating findings into clinical practice is extremely difficult.
There is also a significant gender and cancer-type imbalance: most research involves men with prostate cancer. Women with ovarian, endometrial, anal, or other gynecological cancers are dramatically under-represented, raising equity concerns about which populations will benefit from AI-driven sexual health tools as they move toward clinical adoption.
Despite current limitations, the review authors outline a clear pathway to clinical integration. Validated AI tools could support early risk stratification of sexual dysfunction, trigger timely referrals to sexual health specialists, and be embedded in electronic health records to prompt routine discussions that are currently avoided due to time constraints or discomfort.
In radiation oncology, AI-based autosegmentation of sexual function structures like neurovascular bundles and the internal pudendal artery could be integrated into routine treatment planning, reducing inter-clinician variability and ensuring these structures are consistently spared when feasible.
NLP tools that extract sexual health symptoms from unstructured clinical notes could enable real-time longitudinal monitoring, identifying patients who are silently experiencing problems but have not been formally assessed or referred for support.
Realizing these benefits requires higher-quality studies with external validation, calibration testing, and transparent reporting. Multidisciplinary guidelines are needed to standardize AI tool use across cancer types, clinical settings, and patient populations, ensuring that benefits reach all patients equitably, not just those treated in high-volume academic prostate cancer centers.