Aging is a complex and personal, cumulative, and irreversible phenomenon that goes well beyond chronological age [1]. It involves several biological events associated with a great variety of molecular and cellular damage, leading to the gradual loss of physiological and immunological reserves and a greater risk for neoplasia-related death [1,2]. Assuming that the elderly population is heterogeneous, this population must be considered not only concerning their chronological age. Thus, an objective analysis of their living conditions as well as aspects related to oncological disease and its therapy is also required [3].

The International Society of Geriatric Oncology has recommended the use of the Comprehensive Geriatric Assessment (CGA) for the evaluation and follow-up of elderly cancer patients [4]. The CGA is a multidisciplinary tool that uses validated instruments to evaluate several elderly health condition domains, such as functional, cognitive, psychological, social, clinical, and nutritional aspects, as well as comorbidities and the use of medication, among others [5,6]. It is also strongly recommended by the geriatrics and gerontology fields in general because it is, in a complex and heterogeneous context, an objective, measurable, and reproducible form of evaluation, adding possibilities to standard clinical laboratory evaluations [7,8]. However, there is no consensus about what and how many instruments should be used. Employing CGA in practice, however, has become a huge challenge, and owing to its complexity and time spent in its application, it is often underutilized by oncologists and not judged as a completely satisfactory solution in practice, which has served as a stimulus for the construction of simpler tools that have the power to predict outcomes and guide clinical decisions [5,9].

The accurate prediction of a disease outcome is one of the most interesting and challenging tasks for physicians. As a result, a growing trend was noted in the studies published during the past years that applied machine learning (ML) algorithms for modeling cancer survival. This type of algorithms can discover and identify patterns and relationships between them, from complex databases, while they are able to effectively predict future outcomes of a cancer type [10]. On the basis of the study by Kourou et al [11], the accuracy of cancer prediction outcome has significantly improved by 15% to 20% in the previous years, with the application of ML techniques.

A study combining data from 4 cohorts involving the elderly, 1 including elderly people with neoplasms, proposed to explore the performance of various ML classifiers (Naive Bayes [NB], k-nearest neighbors, artificial neural networks, random forest, and logistic regression) regarding death prediction in 6 months [12]. Another study used ML to predict mortality of patients in 3 to 12 months and to identify patients who could benefit from palliative care [13]. However, no ML application has been proposed using CGA to classify elderly cancer patients.

Thus, the primary aim of this study was to propose and develop predictive models, using ML and CGA, to estimate the risk of early death in elderly cancer patients. The secondary aims were to optimize the CGA through the selection of the most appropriate instruments.

The ability of ML techniques to predict early mortality in a heterogeneous cohort was tested in 608 elderly cancer patients (aged over 60 years), admitted to the oncogeriatrics sector of the Instituto de Medicina Integral Prof. Fernando Figueira - IMIP, from January 2015 to July 2016. The IMIP is a teaching hospital and cancer center located in Recife, Pernambuco, Brazil. On admission to the cohort database, the patients were evaluated by CGA questionnaires presented in Table 1. The questionnaires were collected by a multiprofessional team, comprising a clinical oncologist, a geriatrician, a physiotherapist, a physical educator, a speech therapist, an occupational therapist, and a nutritionist. The project was approved by the IMIP Ethics Committee on Human Research on June 30, 2016, under number 58298316.5.0000.5201.

The first step was to remove patients presenting redundancies and/or incomplete questionnaires/features. A total of 543 patients remained after that. Data normalization technique for equalizing the range of features, usually employed in the database before feature selection and learning phase, is of important concern in pattern recognition and computer-aided diagnosis [23]. The most common normalization method used during data transformation is the min-max (where the features are mapped into a predefined range, varying from 0 or −1 to 1). The main advantage of min-max normalization is that it preserves the relationships among the original data values [24]. In this work, all features were normalized in a (0,1) interval, calculated as in equation, where v′ is the value normalized, v is the original value, v_min is the minimum value of corresponding feature, and v_max is the maximum value of corresponding feature: v’=(v−v_min)/(v_max−v_min).

Predictive modeling is the general concept of building a model that can make predictions. Typically, such a model includes an ML algorithm that learns certain properties from a database to make those predictions. We have presented below a brief summary of the commonly used supervised learning algorithms:

The purpose of this work was not to introduce the highest accuracy prediction model. The goal was to designate the most relevant questionnaires to evaluate elderly health condition domains in CGA. Therefore, in the experiments, we always used the same configuration with the default parameter values in Weka (Waikato Environment of Knowledge Analysis) from The University of Waikato, version 3.8.3. The advantage of using default parameters is that it does not introduce optimistic bias by tuning the parameter to maximize performance on the test data. Figure 1 shows more details about the values used in each predictive model.

Cross-validation (CV) [28] is one of the most widely used methods to assess the generalizability of predictive models [29] and is subject to ongoing active research [30]. K-fold CV comprises dividing the database into K parts (folds) of equal sizes. For this study, a 10-fold CV is used, and each part is held out in turn and the predictive model (J48, MLP, or NB) is trained on the remaining nine-tenths; then, its error rate is calculated on the holdout set. Thus, the learning procedure is executed a total of K times on different training sets (each of which have much in common). Finally, the K error estimates are averaged to yield an overall error estimate. In this work, the folds are made by preserving the percentage of samples for each class.

The learning procedure and the subsequent prediction of predictive models can be affected by the problem of imbalanced database [31]. The balancing issue corresponds to the difference in the number of samples in the different classes. The resulting database presented 92 deaths within 6 months of admission to the service and 451 patients alive at the end of that period. All deaths were attributed to cancer (treatment complications or disease progression). With a greater imbalanced ratio, the decision function favors the class with the largest number of samples, usually referred as the majority class. The way to fight this issue was to generate new training sets on 10-fold CV by random sampling so that the proportion between classes remained at one-to-one.

The area under receiver operating characteristics curve, or simply area under curve (AUC), has recently been proposed as an alternative single-number measure for evaluating the generalization of learning algorithms [32]. This measure is far better than classification accuracies when the 2 classes are unbalanced and the cost of misclassification is unspecified [33]. An area of 1.0 represents a model that made all predictions perfectly, and an area of 0.5 represents a model as good as random. AUC can be broken down into sensitivity and specificity:

Feature selection is an important and frequently used technique for dimension reduction by removing irrelevant and/or redundant information from the database to obtain an optimal feature subset. A 10-fold CV was used to evaluate all possible combinations of CGA questionnaires, presented in Table 2, to estimate the risk of early death in elderly cancer patients. Thus, in each fold, the combination of questionnaires with highest AUC is selected. The same folds are applied to all 511 combinations. Tiebreaking is handled by choosing the smallest set of questionnaires. The occurrence of questionnaires selected on the 10-fold CV, using predictive models, is presented in Table 2. In Figure 2, the flowchart of our methodology is shown.

Tables 3-5 show the sensibility, specificity, and AUC values expressed as mean (SD) on the 10-fold CV for the NB, J48, and MLP. The subsets of CGA questionnaires, presented in these tables, consider the occurrences of Table 2. The subset of questionnaires with occurrence ≥0, for example, uses all set of CGA questionnaires, as it considers all occurrences greater than or equal to 0. The other subsets use the same logic and are detailed in the footnotes under the tables. In each metric, according to the paired t test, the P value is calculated considering the subset of questionnaires with occurrence ≥0. The experimental results demonstrate that the feature selection can discard questionnaires and finally find out subsets that reduce the dimensionality of data to make the predictive models more efficient and the results more accurate. Thus, a simplified predictive model aiming to estimate the risk of early death in elderly cancer patients is proposed herein, minimally composed by the Mini Nutritional Assessment-Short Form (MNA-SF), accompanied or not by the Karnofsky performance scale (KPS) and/or the Mini-Mental State Examination.

Results indicate that the MNA-SF has greater predictive power to estimate the risk of early death in elderly cancer patients as it was selected on the 10-folds. MNA-SF is a rapid test validated for screening for nutritional risk and malnutrition in the elderly population. The predictive value of MNA-SF for early death may be related to the fact that the 6 MNA-SF questions cover areas other than just nutrition, which are frequently included in the CGA, such as mobility, neuropsychological disorders, and self-reported health, in addition to nutrition aspects, including weight loss, reduced food intake, and body mass index. In fact, low MNA-SF may reveal the effects of advanced disease in the overall health of patients, which also affects cancer-related mortality. A Brazilian study showed that abnormal nutritional status was an independent factor associated with hospital death among older patients with various chronic diseases, including cancer [34]. A similar association was also demonstrated in elderly Asian cancer patients who would receive first-line chemotherapy [35]. Finally, a French multicenter study with 348 elderly cancer patients aged 70 years and above also found that low MNA scores were associated with increased risk of premature death [36].

The results also indicated that KPS questionnaire has proven itself a valuable tool to estimate the risk of early death in elderly cancer patients. In the past decades, various studies have demonstrated the prognostic value of the KPS not only primarily for various cancers [37-40] but also for other disease entities [41]. It can also be considered as a significant indicator of hospitalization and survival time, in addition to identifying risk groups to assist in the orientation of patients to geriatric outpatients [42].

The efforts of this paper are a starting point. They provide solid evidences and some clinical recommendations. We proposed and developed simple ML models for the prediction of early death in elderly cancer patients. These models are accurate and precise and could be possibly used by clinicians to make proper treatment plans. However, additional research is needed to continue to strengthen the evidence base.

The results showed that the MNA-SF and KPS have the highest predictive power to identify elderly patients at risk for early death. We strongly recommend that these questionnaires be incorporated into regular geriatric assessment of older patients with cancer.

The MNA-SF and the KPS requires only a few minutes to be completed. In addition, both can be easily managed by any member of the multidisciplinary team to help in the early identification of patients at risk, providing information that assists in the planning of interventions and improving the adherence to CGA in daily clinical oncology practice.

This study also has limitations that should be considered. This is a nonrandomized, single-center, exploratory study of a heterogeneous patient population similar to a real-life population of older patients with cancer. Conversely, some of its weaknesses could be considered the main strengths of the study: this is one of the few studies in Brazil that, in the clinical practice context of a Unified Health System oncology unit, investigated the use of ML algorithms in the prediction of early death in elderly cancer patients.