Melanoma, the most aggressive form of skin cancer, poses a significant health challenge worldwide [18]. Among the various types of melanoma, lower limb melanoma accounts for a substantial proportion of cases [19]. Accurate prognosis and survival estimation are crucial for guiding treatment decisions and improving patient outcomes [20]. Deep learning techniques, such as DeepSurv, have shown promise in predicting patient survival rates based on clinical and genetic features.

This study explores the potential of DeepSurv in advancing our understanding of lower limb melanoma and its implications for personalized medicine. Deep learning model specifically designed for survival analysis [4]. Unlike traditional statistical methods, which assume cox proportional hazards, DeepSurv is capable of handling complex, high-dimensional data and non-linear relationships [21]. It predicts survival probabilities over time, enabling precise and dynamic risk assessments for patients. It also can leverage diverse data sources, including clinical records, histopathological data, and genetic profiles. Integrating this information can provide a more comprehensive picture of the patient’s condition and potential risk factors [22, 23].

Traditional prognostic models in melanoma often rely on a limited set of variables, leading to generalized estimates [7, 24]. DeepSurv, with its ability to capture complex relationships, may offer more precise and individualized survival predictions for patients, helping oncologists tailor treatments to specific needs.

Within the CoxPH model, a variety of factors such as age, sex, AJCC, surgical interventions, chemotherapy, and Reg_LN_Sur were identified as significant risk factors impacting the domain of LLM. Furthermore, the CoxPH model exhibited good C-index, attesting to its commendable predictive precision.

DeepSurv’s capacity to analyze vast amounts of data may lead to the discovery of novel prognostic markers for lower limb melanoma. These markers could unlock new avenues for targeted therapies and early intervention strategies. And it also accounts for the timing of events, providing time-dependent survival probabilities. This capability is particularly relevant in melanoma, where the disease progression can vary over time.

So, the newly developed DeepSurv model, consisting of an intricate neural network with multiple discerning layers, achieved remarkable performance with a higher C-index, which is 0.852. Notably, there was a noticeable disparity in the calibration curves between the DeepSurv and CoxPH models. The DeepSurv model demonstrated a more evenly distributed profile, aligned harmoniously with the leading-diagonal line, which indicated its superiority. This superiority was further evident in the AUC curve, where the DeepSurv model exhibited exceptional smoothness that surpassed its CoxPH counterpart, reaffirming its prowess in predicting 3-, 5-, and 8-year mortality and survival-time outcomes for patients with LLM [16]. The reason we choose 3-, 5-, and 8-year mortality is due to previous research. Lower limb melanoma survival may have identified these time points as important for assessing long-term outcomes or for making comparisons with other studies [25, 26]. Also, in cancer stat facts of melanoma (https://seer.cancer.gov/statfacts/html/melan.html) the 5-year survival is 93.5%. Therefore, we collected data and made observations both before and after the 5 years survival.

The correlations of our deep learning results with those of other authors in the field provides essential validation and contextualization of our findings [22, 23, 27, 28]. It allows us to gauge the generalizability and clinical utility of our model, identify potential challenges, and highlight its strengths in specific patient populations or cancer types. By embracing collaboration and comparison across studies, we can collectively advance the field of deep learning for cancer and foster its seamless integration into clinical practice for improved patient outcomes.

The DeepSurv model’s predictions have several values in healthcare application. For high-risk patients’ intensive treatments or closer monitoring should be able identified by the doctors which will give to customize treatment plans, while low-risk patients be able to use fewer intensive treatments, for a better cause. Also, in clinical daywork we face limited resources, with DeepSurv model’s we can allocate the appropriate amount resources for whom needs it, this way we can raise the efficiency. Make sure high-risk patients would receive appropriate follow-up and specialized care so clinicians can offer patients with precise information, facilitating the right prognosis and treatment plans. Moreover, for patients with better outcomes, the DeepSurv model’ can provide with more accurate long-term care plans, with monitoring the recurrence, and other life matter concerns. Based on the result of each model, for cancer epidemiology, trends, and survival outcomes we still can use SEER database to give us the insight based on what SEER database has.

This study has a number of restrictions, which should be acknowledged. First off, the SEER database’s lack of essential prognostic factors, such as complex surgical procedures, specialized radiotherapy protocols, precise treatment with chemotherapy regimens, pharmacological interventions, and related information, limited the breadth of our findings regarding patients with LLM. Second, because the dataset only included data from a few US states, the generalizability of our study findings was constrained by the lack of external validation. The DeepSurv model will be improved in the future by adding more, more varied information with a wider geographic reach. Thirdly, the DeepSurv model’s hidden layer’s intrinsic opacity, which functions as a computational “black box,” made it difficult to understand the specific mechanics underlying its ability to forecast the future and the decision-making process it uses. Through thorough study and clarification, we aim to address the aforementioned constraints in our next study [29].