1Clinical Dermatology, IRCCS Istituto Ortopedico Galeazzi, 2Department of Biomedical, Surgical and Dental Sciences, University of Milan, Milan, 3PhD Degree Program in Pharmacological Sciences, Department of Pharmaceutical and Pharmacological Sciences, University of Padua, Padua, 4Department of Cardiological, Vascular, and Thoracic Sciences, and Public Health, University of Padua, Via Loredan 18, IT-35128 Padova, 5Villa Santa Maria Institute, Neuropsychiatric Rehabilitation Center, Tavernerio (Como), 6Veneto Tumor Registry, Azienda Zero, 7Surgical Oncology Unit, Veneto Institute of Oncology IOV-IRCCS, 8Veneto Regional Authority, Padua, 9Fondazione Policlinico Universitario A. Gemelli IRCCS, Rome, and 10Università Cattolica del Sacro Cuore, Rome, Italy. *E-mail: alessandra.buja@unipd.it
Accepted Oct 28, 2020; Epub ahead of print Nov 2, 2020
Acta Derm Venereol 2020; 100: adv00323.
doi: 10.2340/00015555-3680
The incidence of cutaneous malignant melanoma (CMM) in Italy has increased in the last decade, leading to public health concern and rising costs of healthcare (1, 2). In addition to individual susceptibility to development of CMM, several environmental variables influence prognosis in this disease. These variables include social disparities, socioeconomic status, education and marital status (3). However, the impact of these variables on costs is unknown. The current study used a new methodology, based on an artificial neural network (ANN), to decodify this complexity by simultaneously describing the relationships between clinical, sociodemographic, outcome, and cost variables, and grouping patients into clusters (4, 5).
This study evaluated a collaborative registry of 556 patients (Veneto Tumor Registry & Veneto Oncology Network)1, who were diagnosed with CMM by a board certified dermatopathologist in 2015 in 4 of the 7 provinces of the Veneto Region in Northern Italy (3). For each patient, the CMM registry includes a set of tumour characteristics, including: tumour-node-metastasis (TNM) stage at diagnosis; Breslow thickness (mm); Clark’s level of invasion (I–V); presence of ulceration (yes/no); site (trunk, head, limbs); cost categories tertiles (scintigraphic, surgical, medical, instrumental, cyto/histological, microbiological, blood examinations, radiotherapeutic, radiological and total (costs based on “Hospital Discharge Forms” (SDO)) CMM-specific mortality.
Costs were assessed from the perspective of the Italian National Health System (Italian NHS), taking only direct costs into account. Patients were linked via unique anonymous identification codes to all administrative data regarding their hospital admissions, day hospital service use, drug usage, visits to emergency services, medical devices used at home, and hospice admissions. These data were used to compute the direct costs for each patient in the 5 years after diagnoses of their CMM.
Descriptive analyses were performed using absolute and relative frequencies for categorical variables. A semantic connectivity map was constructed using Auto Contractive Map (AutoCM, Semeion©, Rome, Italy) to elucidate variable links (4). The system highlights the natural links on a graph based and distances between variables reflect the weights of the ANN (Appendix S1). AutoCM has many relevant features: (i) non-linear associations among variables are preserved; (ii) patterns of connections between clusters of variables are captured; and (iii) complex similarities among variables emerge.
Clinical/histological and demographic data for the 556 included CMM are summarized in Table SI.
The AutoCM results are shown in Fig. 1. The item “1–6 mitoses” is the centre (main attractor) of our unsupervised analysis, demonstrating its clustering value vis-à-vis the spread of its 4 main branches (strength > 0.60) depicting 4 endotypes. The use of radiotherapy, education and marital status were central descriptors in our database. The 1st endotype grouped together (or “clusters”) those patients with advance-stage CMM who had nodular and ulcerated CMM, a high risk of death, and a heavy economic burden. The 2nd endotype clustered patients > 60 years of age who had CMM on the trunk or face, and high procedural and therapeutic costs. The 3rd endotype clustered patients with stage Ib CMM. The 4th endotype clustered patients with no radiotherapy costs, comprising 4 main subsets, each with their own biological and socioeconomic variable items.
Further details of the items interactions are shown in Appendix S2. Our results, in-line with the idea of precision medicine, also suggest a potential endotype-guided treatment Appendix S3 that may implement CMM follow-up visits.
Fig. 1. Auto Contractive Map (AutoCM) semantic map, showing cutaneous malignant melanoma (CMM) endotypes created by clustering variables belonging to different fields (clinical, therapeutic, histological, demographic and costs). “0_MM”: absence of metastases; “1_MM”: metastases clinically and/or radiologically evident; Breslow_1: <0.76 mm; “Breslow_2”: 0.76–1.75 mm; “Breslow_3”: >1.75 mm; “Breslow_4”: missing data; “Breslow_5”: regression or metastases; “CMM”: cutaneous malignant melanoma; “Death”: melanoma-specific mortality; No diagnosis: no comorbidities. Note: it was opted to exclude branches radiating from “1–6 mitoses” (which has a bond strength <0.60) as endotypes, so the branches “separated” and “high costs of medical therapies” were excluded. Four main endotypes were identified (I, II, III, IV). The fourth endotype has 4 sub-endotypes, which were termed IV-a, IV-b, IV-c, IV-d. The endotypes are as follows: (I) advanced-stage patients: with nodular and ulcerated CMM on lower limbs (Clark IV, Breslow III–IV, TNM IV or II with metastases and >6 mitoses), high risk of death, and heavy economic burden; (II) patients >60 years, widowed, with primary school diploma, with unstable chronic disease, non-nodular CMM on the trunk, or lentigo maligna on the face, with 1–6 mitoses, associated with high costs of SDO (Hospital Discharge Form), diagnostic procedures (visits, cytohistology, radiological and instrumental investigations), and therapies (medical, surgical); (III) patients with stage Ib disease: patients with TNM Ib, and Breslow II with 1–6 mitoses; (IV) a cluster around the absence of costs for radiotherapy, which includes 4 main subsets, each with their own biological and socioeconomic items: (IV-a) female under 30 years old with non-metastatic, regressive and superficial spreading CMM of the upper limb, with 1–6 mitoses; (IV-b) male aged 50–60 years with CMM with 1–6 mitoses, and without microbiological costs; (IV-c) male aged 40–50 years with a high-school diploma, with TNM Ia, Clark II, Breslow I CMM, with no mitosis, associated with moderate radiology costs; (IV-d) a miscellany of CMM patients with 1–6 mitoses, further classifiable as: (a) married and uneducated, with a stable chronic condition, which was associated with moderate costs of cytohistology, surgery, low costs of instrumental investigations, and low total costs; (b) single with a university degree, associated with moderate hospital discharge records (SDO), and low costs of surgery; (c) male aged 30–40 years with acral lentiginous CMM, with 1–6 mitoses, associated with moderate costs of blood tests, specialist visits, and microbiological tests.
Through a combined analysis of the clinical, sociodemographic and economic variables associated with CMM, using an ANN, endotypes were identified that can be used to estimate an individual patient’s final costs based on their baseline characteristics. Based on these patient clusters, further tests can be suggested to add to the routine tests used for patients with CMM at each TNM stage.
Use of a machine learning approach such as this can generate a comprehensive model that establishes complex links between clinical, sociodemographic, outcome, and cost variables. Machine learning has also been used by Finlayson et al. (6) on a small database containing the clinical, therapeutic and molecular features of 237 cases of advanced CMM, with the aim of helping clinicians predict patients survival and make therapeutic decisions.
In our database the main variable, also regarded as the “attractor”, was “1–6 mitoses. Mitotic index is not included in the 8th edition of the American Joint Committee on Cancer TNM system, suggesting that it is optional to report this variable for prognostic purposes. The value of the mitotic index as a prognostic indicator in CMM has been debated, partly because of its moderate interobserver variability (7). The findings of the current study suggest that the mitotic index is relevant for predicting the costs of CMM from baseline information.
The semantic map used in the current study revealed 4 endotypes, through using the database as “learning material” for the evolutive algorithm of the ANN (8). The internal validity of the algorithm is high and measurable; however, further study is required to determine its external validity, since the ANN was designed to find connections among variables in the database in the current study. Evidence to support external validity come from the literature, specifically regarding marital status and education. In line with literature (3), we confirmed that marital status represents a valuable information to be recorded in CMM patients. An association between married status and lower costs may stem from the presence of a partner resulting in earlier diagnosis of a skin cancer. This suggests that patients with little or no formal education are less inclined to access healthcare services (3).
This study has some limitations. Data regarding the patients’ self-reported educational level may not be reliable, although a previous Italian study on the validity of data on education levels recorded in hospital discharge records found that it was in the good-to-excellent range (9). The current study lacks information on patients’ lifestyles and other socioeconomic parameters that might reduce the influence of educational level on the natural history of their CMM.
A strength of this study is its population-based dimension, which minimizes selection bias by using independently- acquired administrative data. The socioeconomic impact of education level and/or income is likely to be mitigated in a universalistic health system like the Italian NHS.
In conclusion, in CMM, clinical variables together with costs were indispensable to cluster patients in endotypes by ANN. Endotypes-guided management is affirming as new promising strategy to guide medical and surgical therapies (10, 11).
IRB approval. Veneto Oncological Institute Ethics Committee (n° 695/20.10.2016).
The authors have no conflicts of interest to declare.
Click to show fullsize
Click to show fullsize