Cutaneous melanoma is the most prevalent etiology of skin cancer. In 2021, approximately 101,280 new melanoma cases were predicted in the United States (1, 2). Currently, innovative approaches such as target therapies and immunotherapies have been introduced in clinical practice for the treatment of cutaneous melanoma (3). More in detail, target therapies are based on the use of drugs targeting specific genetic alterations in candidate genes, blocking specific pathways implicated in the oncogenesis of melanoma (3). The most important is the MAP-Kinase (B-RAF/MEK/ERK) pathway (1). Among immunotherapies, checkpoint inhibitors target specific receptors on T-lymphocytes, such as anti-CTL-A4, anti-PD-1/PDL-1, and anti-LAG-3. They all have improved the survival of melanoma patients in both metastatic and adjuvant setting (4, 5). Therefore, we noticed an advancement in the treatment of cancer over the past decade, albeit it could result in an impaired health related quality of life (HRQoL) (6–8).

In this context, beyond the cancer itself, patients with melanoma should undergo follow-up programs and be aware of their condition in order to improve well-being and HRQoL (4, 9). HRQoL has been defined as the ‘perception of the effects of illness and treatment on the physical, psychological, and social aspects of life’ and it is becoming increasingly recognized as an important therapeutic outcome in cancer therapy. Patients seeking medical advice concerning pigmented lesions initially report a good HRQoL, that could often lower by the time of melanoma diagnosis and following treatments (10).

The female sex also predicted greater improvement in HRQoL over time and younger patients (<50 years old) demonstrated poorer HRQoL (11). Moreover, lymphedema is a significant health issue for cancer survivors as a condition that can severely affect the HRQoL of patients (12–14). Sentinel lymph node biopsy (SLNB) has become the procedure of choice for assessing the lymph node status, designed to be less invasive and, hence, less prone to induce complications than a complete lymph node dissection (15, 16). When necessary, lymph node dissections for melanoma treatment is led to rates between 15.7 and 64.3% of secondary lymphedema (5).

Several generic- and specific HRQoL questionnaires have been developed or are under development (6) and are varying in their sophistication, reliability, and validity (17, 18). Validity problems (mainly concurrent and criterion validity) are evident when comparing HRQoL-results between studies, as there is no consensus concerning the definition of the subjective and multidimensional concept of HRQoL (19).

Among traditional statistical methods, logistic regression is the most common when dealing with binary outcomes, like the presence or absence of a variable (20). Conversely, machine learning is a branch of artificial intelligence centered on algorithms which do not need explicit prior programming to function but automatically learn from available data, creating decision models to complete tasks. Random forest regression (RFR) belongs to the branch of the same group of decision trees of machine learning methods, and it might become an essential part of every step of oncological screening strategies and management of patients, thus leading to precision medicine (21).

Therefore, the present study aimed to identify the weight of factors via a machine learning approach in the perception of the HRQoL in a cohort of patients affected by melanoma.

The final study cohort consisted of 203 patients with melanoma, mean aged 59.25 ± 15.1 years: 56 (27.6%) at the upper limbs and 147 (72.4%) affecting the trunk. A full listing of patient clinical characteristics is listed in Table 1 .

The model, built on 142 patients with no missing variable values and relative SF-12 scores, generated 92 trees, contemplating 44 subjects in testing and 36 subjects in validation of machine learning model. So, the MSE value stood at 0.45 with an R2 of 0.78.

The evaluation of the importance of the variables reported that the site of lesion was the most influential variable on the quality of life based on the decrease of Gini impurity when choosing a variable at the junction of a node in the forest generation, demonstrated the importance of the variable as expressed in Table 2 .

Given these results, an analysis of the out-of-bag performance of the model assured that as the amount of random trees generation in the forest increased, the curves of the training and validation dataset appear and stand below the MSE less than 0.5, validating the robustness of the prediction. Furthermore, we evaluated the correlation between the predicted values in the model and the actually calculated values. We reported a linear regression index of 0.78, considering a good reliability of the prediction, as depicted in Figure 1 .

Given the importance of the site variable in the overall analysis, we divided the sample into two subgroups. The UL random forest model, built on 51 patients with no missing variable values and relative SF-12 scores, generated 99 trees, contemplating 12 subjects in testing and 12 subjects in machine learning model validation. Thus, the MSE value stood at 0.49 with an R2 of 0.53.

The importance of the variables assessment regarding UL melanoma showed that lymphedema become the most influential variable on the PCS-SF-12 scores as it increases the precision of the model in purity of the node. On the other hand, the importance of the variable SLNB remains decisive as expressed in Table 3 .

Given these results, the out-of-bag performance reported that as the amount of random trees generation in the forest increased, the curves of the training and validation dataset reported an error of 0.87, verifying a minor robustness of the prediction. Furthermore, we evaluated the correlation between the predicted values in the model and the calculated values, reporting a linear regression index of 0.53 with a moderate reliability of the prediction.

Concerning the model on trunk melanoma, built on 91 patients with no missing variable values and relative SF-12 scores, generated 63 trees, contemplating as data split 21 subjects in testing and 16 subjects in validation of machine learning model. The MSE value stood at 0.49 with an R2 of 0.67.

The evaluation of the importance of the variables reported that BMI is the most influential variable on the quality of life as it decreases the precision of the model when the variable is excluded. Similarly, the measure based on the decrease of Gini impurity when choosing a variable at the junction of a node in the forest generation, demonstrated the importance of the variable BMI and SLNB as expressed in Table 4 .

The analysis of the out-of-bag performance of the learning demonstrated that with increase of random trees generation in the forest, the curves of the training and validation dataset appear and stand with an OBB error of 0.49, verifying the robustness of the forecast. Furthermore, we evaluated the correlation between the predicted values in the model and the calculated values. We reported a linear regression index of 0.67, considering a good reliability of the prediction.

In summary, as shown in Figure 2 , patients in overall model are affected in HRQoL by the melanoma site and SLNB; on the other hand, the HRQoL of patients with UL melanoma is predominantly influenced by the lymphedema, while the HRQoL of trunk melanoma patients is affected by the weight of the BMI. Intriguingly, positivity of SLNB and age did not seem to influence PCS-SF-12 in a decisive way.

This study aimed to identify the weight of predictors of obtaining PCS-SF-12 in a cohort of people with melanoma through a machine learning approach. The main results reported that the degree of lymphedema and BMI could predict the weight in the perception of HRQoL in patients with upper limb and trunk melanoma, respectively. A melanoma treatment factor such as SLNB influences the quality of life, but not its analysis and positivity, and also age variable which does not substantially affect the machine learning model. Risks of SLNB are low but include a risk for development of lymphedema, although this risk is higher in the lower extremity than in the upper one (33).

Lymphedema is one of the more significant complications that can occur after a lymph node dissection of the axillary or inguinofemoral lymph nodes (34). Usually, these lymph node dissection schemes are performed in the context of node-positive breast cancer and melanoma (35). Secondary lymphedema is a frequent complication after lymphadenectomy in melanoma patients, although few studies in melanoma population adequately characterize risk factors for lymphedema; in addition the sample size is frequently limited (36). Several factors have been implicated in the modulation of lymphatic transport, namely, obesity, age, anatomic location, strength in the contraction of surrounding muscles, damage to sensory nerves, lymphatic leak requiring prolonged drains, hematoma, and wound infections delaying adjuvant treatments (37). Therefore, obesity has demonstrated in preclinical models and in patients to cause decreased lymphatic pumping and leakiness of lymphatics (38, 39).

Indeed, there is a complex relationship between obesity, melanoma, and lymphedema. Preclinical and clinical data showed a correlation between obesity and Breslow thickness, suggesting a higher biologic aggressiveness of melanoma in obese patients (40). Moreover, a recent study has showed that increasing BMI, specifically morbid obesity, is associated with an increased risk of wound complications, namely, lymphedema in the setting of both axillary and inguinal lymphadenectomy (41). In addition, pain at the site of lymph node surgery, and cellulitis of the limb certainly have been reported to increase the risk of lymphedema (42). Given the high number of patients developing lymphedema after inguinal lymph node dissection (ILND), the impact of these chronic issues related to the quality of life and infection on the health care system is not trivial (43, 44). A greater number of lower limb lymphedema symptoms and higher severity may catalyze those negative effects, indicating that proper management of modifiable factors may have a positive impact on patient health and well-being (36). Ryan et al. described the negative experiences of patients with the frustration of symptoms and treatment, diminished self-confidence and poor self-perception, and changes to body image. These findings were further supported by experiences of isolation, depressive thoughts, and fears of the future of lymphedema patients, all of which are implicated in patient HRQoL and psychosocial well-being (45–48). Therefore, Gjourp et al. affirmed that the negative impact of melanoma-related limb lymphoedema on HRQoL was significant in multiple domains of HRQoL (19, 49).

The random forest approach is a machine learning algorithm based on several decision trees, proving its success in both regression and classification problems in recent years and is one of the best machine learning algorithms used in many different fields (50). Developed in the 1990s, random forests have become known for their state-of-the-art capability in classification or regression, and their ability to handle categorical or continuous variables, and in dealing with missing data (51, 52). In addition, in most implementations, the so-called out-of-bag or generalization errors are automatically calculated and their performance is not particularly sensitive to the few hyperparameters that are required to tune the models. Consequently, the popularity of these models in the process industries is growing rapidly, with applications, for example, in predictive modeling.

This study is not free from limitations. Indeed, RFR provides a measure of variable importance, but a current limitation is that no systematic method exists to estimate the shared variances of the variables. Moreover, the mean age of participants included in our cohort was 59.2 years, representing a relatively older cohort of survivors. Although this is representative of the population of people typically diagnosed with cancer, without younger survivors included, the full spectrum of survivorship experiences is not identified. At least, we have included only patients with upper limb and trunk melanoma.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving human participants were reviewed and approved by the Istituto Nazionale Tumori Pascale, Naples, Italy. The patients/participants provided their written informed consent to participate in this study.

Study conceptualization: MP and AdS. Methodology: NM and AdS. Investigation, MP, CC, and ES. Resources: MP and AA. Formal analysis: NM. Data curation: MP and AdS. Writing—original draft preparation: MP and NM. Writing—review and editing: AdS. Visualization: CC, ES, and AA. Supervision, AA and AdS. Submission: AdS. All authors listed have made a substantial, direct, and intellectual contribution to the work and approved it for publication.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.