Background

Amyotrophic lateral sclerosis is a disease characterized by the progressive and irreversible degeneration of motor neurons, which causes deficits in the ability to control movement, breathing, and, in 50% of cases, in cognitive and behavioral functioning [1,2,3]. The cause of ALS is still unknown and there is no treatment to cure it. Hence, there are only alternatives of palliative care and medication to delay the progress of the disease [4, 5]. Diagnosing patients with ALS represents a challenging task due to its complex pathogenesis and the absence of specific biomarkers [6, 7]. The diagnosis is based on clinical presentation, progression of symptoms, and the exclusion of other diseases supported by tests such as Electromyography (EMG). Such a process requires an average of 10–18 months from the onset of symptoms to confirmation [8,9,10,11]. What is more, the diagnosis is considered slow and late given the characteristics of ALS, in which life expectancy after confirmation is of 2–5 years [12].

Despite being described by Jean-Martin Charcot more than 100 years ago [13], ALS is considered a rare disease, and, to this date, there are not many countries with records of epidemiological data. In a few European countries, as well as in the United States, epidemiological records show that the incidence rate of ALS is of 1–2 cases per 100,000 individuals per year, while the prevalence is approximately 5 cases per 100,000 individuals, which for van Es et al. [3] reflects the fast lethality of the disease. A worldwide increase in the number of ALS-affected individuals is expected, rising from 222,801 cases in 2015 to 376,674 by 2040, according to the projection made by Arthur et al. [14]. The aging of populations and the consequent rise in the number of individuals within the age group with a more considerable risk for ALS, which is of 60–79 years, represent the probable culprits for the 69% worldwide increase [14].

Considering the intrinsic aspects of ALS, it is critical to promptly search for diagnostic support systems, as well as for alternatives that intermediate essential communication, autonomy, and promote quality of life to patients. From this standpoint, several technology-based studies have been developed. These investigations typically provide auxiliary resources for diverse aspects regarding ALS, going from what pertains to patients and their caregivers to matters related to outpatient care in organizational health entities [15,16,17].

Technologies developed for health encompass and collaborate in positive progressions in remarkable ways, such as with the diagnosis of ALS [18, 19], monitoring of disease progression [20], monitoring of food intake [21], communication intermediation [22,23,24,25], autonomy [26], and other applications based in artificial intelligence, as it has been reviewed by Schwalbe and Wahl [27]. Automated systems for disease diagnosis, for instance, are computational tools composed of ML techniques that, based on the processing of biomedical signals, are capable of aiding the detection of neuromuscular disorders [28]. These systems contain expert information of specific domains, which provide health professionals with decision-making support and represent strategies and measures adopted in the care of patients [29].

Recently, in the context of ALS, Grollemund et al. [30] published a comprehensive review that presents and investigates ML models. Thus, it uses or combines different data types from individuals with ALS (clinical, genetic, biological clinical, and imaging), in three-class applications: diagnosis, prognosis, and risk stratification. In conclusion, the authors point to promising advances with this approach in the academic and clinical field in the ALS ecosystem. In this perspective, this SLR complements Grollemund et al. [30] in analyzing ML models in applications for ALS using specifically biomedical signals.

Biomedical signals consist of data from a studied physiological system and their processing aims mainly to extract relevant information [31, 32]. This information can enhance data-driven artificial intelligence techniques, especially ML algorithms, and it is used to support the diagnosis of various diseases [27]. There are several types of biomedical signals, as EMG, electroencephalogram (EEG), electrocardiogram (ECG), electrooculogram (EOG), gait rhythm (GR) and magnetic resonance imaging (MRI). Regarding the ML models, Artificial Neural Network (ANN), decision tree (DT), support vector machine (SVM), and K-Nearest Neighbor (KNN) are particular examples of techniques that have been extensively considered in the healthcare realm, including in the context of ALS [33,34,35,36].

Objective

The chief goal of this systematic literature review is to investigate ML-based approaches, in tandem with the biomedical signals, that contribute to the practical and scientific advancement of aspects in the field of ALS. In this manner, it is expected to provide an overview of the matter at hand, considering the identification of the most-used biomedical signals and ML-based models, in addition to gathering details of primary studies, such as the purpose, the performance of algorithmic models, and experimental data, to identify strengths and opportunities for future researches.

Methods

We have developed this research considering the systematic review guidelines proposed by Kitchenham [37]. In the perspective of investigating technological applications in ALS, this study aims at (i) identifying the most applied biomedical signals; (ii) identifying for what purposes those are used; and (iii) verifying the usage of ML techniques or intelligent approaches to the processing of those signals. Hence, the research questions (RQ) were elaborated on this premise (see Table 1, presented below).

Table 1 Research questions
Full size table

The primary studies searching and screening process in the scientific databases were categorized into four stages, according to what is displayed in Fig. 1. In the first stage, an initial set of articles was selected from the output of searches carried out in the IEEE Xplore, Web of Science, Science Direct, Springer, and PubMed databases. The following search strings (STR) were used in this first stage:

  • STR01: (((“signals processing” OR “signals biomedical”) OR (“smart systems” OR “machine learning” OR “artificial intelligence” OR “computational intelligence” OR “algorithm” OR “algorithms”)) AND (“amyotrophic lateral sclerosis” OR “als”));

  • STR02: (((“signals processing” OR “signals biomedical”) OR (“intelligent systems” OR “machine learning” OR “artificial intelligence” OR “algorithms” OR “Computational Intelligence”)) AND (“amyotrophic lateral sclerosis” OR “als”)).

In the second stage, the predefined inclusion criteria (IC), presented in Table 2, were applied to the initial set of articles from the previous phase. Primarily, an IC delimits the boundaries or scope of the investigation and possibilities the generation of a new subset of papers with a more significant probability of answers to the RQ. In such a context, the subset includes research articles from the last ten years that have been published in journals and are directly related to the principal area of interest of this systematic review.

Fig. 1
figure 1

Methodology steps

Full size image
Table 2 Inclusion criteria
Full size table

In the third stage, after screening the articles through the IC, the verification and removal of duplicate papers were carried out. Besides, a filtering procedure—by considering title, abstract, and keywords—was performed to exclude papers that did not present specific terms related to the theme of this review. Such a process was guided by the exclusion criteria (EC) (see Table 3) and was executed through the Rayyan web application [38].

Table 3 Exclusion criteria
Full size table

In the fourth stage, the total reading of the filtered articles was performed. Hence, it was executed the quality assessment (QA) protocol (see criteria in Table 4). In the QA procedure, each criterion was attributed points measuring the relevance of the article to the target subject of this research. The points were distributed in the form of weights (w), considering suitable responses to the QA criteria, present in the primary studies, with 1.0 being the most relevant weight and 0 the lowest:

$$w_{\text {QA}} = \left\{ \begin{array}{ll} 1.0, & \text {yes},\; \text{fully}\; \text{describes} ,\\ 0.5, & \text {yes}, \; \text {partially} \; \text {describes},\\ 0, & \text {does}\;\text {not}\; \text {describe}.\\ \end{array} \right.$$

A score, the arithmetic mean of the points of the QA criteria (Eq. 1), was generated for each article. In this case, all articles that obtained a score greater than or equal to 0.5 (\(0.5 \le \text{score} \le 1\)) were selected for this research and constitute the final set of articles.

$$\text{score} = \frac{1}{\text {QA}}\sum _{i=1}^{\text {QA}} w_{\text {QA}_{i}}$$
(1)

Records relevant to each stage, as well as the data extracted from the articles, were properly gathered in spreadsheets and the Rayyan web application [38] for data extraction. Data, such as year of publication, authors, and possible responses to the RQ, were extracted from the set of articles of the fourth stage. They permitted the final analysis and fulfillment of the objectives of this systematic review.

Table 4 Quality assessment
Full size table

Results

The results obtained from the searching and screening process of primary studies are synthesized in Fig. 2. In the first stage, 10128 candidate articles were identified after searching with the STRs. In the subsequent phase, three refining procedures based on IC (Table 2) were applied, and 9914 papers were discarded for not meeting the IC. At this point, 214 articles were considered appropriate for inclusion and analysis in the following stage. In stage three, the applied filters, based on the EC (Table 3), removed 186 articles amongst duplicates and those missing the target terms of the search. In this manner, 28 studies were selected for full-text reading and assessment through the QA criteria. After the QA procedure, the fourth and latter phase, 18 papers exceeded the pre-established minimum score, according to the result presented in the respective column in Table 5, and were included for analysis and definitive investigation in this review.

Fig. 2
figure 2

Result of the search and screening process of primary studies for this systematic review

Full size image
Table 5 Set of selected articles and their main characteristics
Full size table

In sum, considering the 18 articles included in this research, the results presented in Fig. 3 evidence three major classes of probable practical applications of biomedical signals processing and machine learning within the context of the ALS disease: diagnosis (or classification), communication, and survival prediction. In addition to categorizing the purposes of such studies, Fig. 3 highlights the number of biomedical signals used and the respective classes that utilized them. Four distinct types of signals were identified: EMG, EEG, GR, and MRI.

Fig. 3
figure 3

Summary of the signals used and their objectives

Full size image

Of the analyzed studies, 44.44% focus on the processing of the EMG signal, the most used biomedical signal (see Fig. 3), and specifically for classification. That is, for the diagnosis of individuals amongst healthy controls (HC), ALS patients, and, in some cases, other diseases (OD). With the same objective, especially for classification, 16.67% of the studies use GR and 11.11% MRI. The MRI signal was additionally used in a particular article for survival prediction of ALS-afflicted individuals, which represents 5.56%. In the communication class, 22.22% of the studies focus exclusively on the approach through the processing of the EEG signal, being this the only one presented for that purpose. This first general analysis of the studies, identifying the purposes of the articles and the signals used, answers research questions RQ01 and RQ02.

Other significant and specific characteristics extracted from the 18 articles included in this study are summarized in Table 5, to support the analysis and answer research questions RQ03, RQ04, and RQ05. Regardless of the classes observed, diagnosis, communication, or survival prediction, all studies used ML algorithms. Alternative algorithmic models were employed and, according to the performance analysis of the algorithm concerning accuracy (Acc), specificity (Spe), or sensitivity (Sen) metrics of evaluation, the best or the only proposed model of each work is shown in Table 5, as well as their respective performances.

For testing, validating, and appraising the proposed approaches in the studies, the algorithmic techniques were applied to a set of data from individuals, distributed in different group combinations of HC, ALS, and/or myopathy, or other neurological diseases. The number of individuals and the type of participating groups in the experiments of each study is specified in Table 5 and summarized in Fig. 4. Moreover, Table 5 describes the source of the dataset and specifies whether they come from public or local repositories.

Fig. 4
figure 4

Number of individuals used in the studies

Full size image

Description of the diagnosis studies

Diagnosis of ALS patients is the most numerous task described among the selected papers, accounting for 72.22% of the studies. Considering only this class, Fig. 5 presents an overview of the number of biomedical signals employed. The use of the EMG signal stands out, being addressed in 61.54% of the studies aimed at diagnosis [39,40,41,42,43,44,45,46]. GR is applied in 23.08% of the studies [54,55], have developed artifacts that promote the communication improvement class of this SLR, exclusively through the EEG biosignal.

In a study conducted by Sorbello et al. [52], a framework was proposed through the operation of a brain–computer interface (BCI) system to control a humanoid robot and promote minimal autonomy to patients with ALS. Generally, the system structure, called brain–computer robotic interface (BCRI), is composed of a BCI system, EEG and eye-tracking devices, and a network system to connect the BCI system to the robotic system. The ML LDA algorithm is used after preprocessing and EEG feature extraction to correctly classify and translate the user action into control commands for the humanoid robot. The authors evaluated the proposal by conducting experiments on four subjects at the HC and four subjects with ALS. The results were satisfactory, and the proposed framework for enabling communication for patients with ALS was validated after all participants were able to control the humanoid robot.

Liu et al. [53] developed an approach by applying the concepts of fractal dimension (FD) and Fisher’s criterion to optimize the selection of EEG channels and the characterization of the data obtained from the signal. In this manner, the authors aimed at improving the classification capacity of an ML algorithm in a BCI system for patients with ALS. Two methods for estimating FD, Grassberger-Procaccia (GPFD) and Higuchi (HFD), were implemented. The key features of 30 EEG channels were extracted and concatenated into a single vector to serve two algorithmic models: KNN and LDA. After tests performed on five subjects with ALS, the results were satisfactory and the GPFD method surpassed the HFD. The performances of the two algorithms, KNN and LDA, were significant and similar, with 95.25% Acc, when compared with the input data containing the 30 EEG channels.

The existence of a simple interface with an accurate and fast information transfer rate is essential to maintain communication efficiency in a BCI system based on EEG signals for people with ALS. For addressing such matter, Mainsah et al. [54] developed a data-driven Bayesian early stop** algorithm, called DS, to optimize the feature selection process of an ERP-based P300 BCI speller, in which ERP stands for event-related potentials. Besides, a variation of the DS is proposed with the application of statistical modeling through Bayesian inference for language predictability, called DSLM. Features correlated with the user’s interest were extracted from the EEG signal to train the stepwise LDA classifier. In the research, the designated online tests were performed with 10 subjects with ALS. Both DS and DSLM algorithms proved to be efficient in minimizing the character selection time and with an average accuracy of 75.40% and 76.39%, respectively. There was no statistical difference between the algorithms.

In the same context of Mainsah et al. [54] and Miao et al. [55] proposed an ERP-based BCI display approach using as strategy a new speller paradigm with peripherally distributed stimuli with the possibility of feedback in the center of the display. The EEG signals were recorded and analyzed using preexisting software from the BCI platform. The features were extracted from data acquired offline from 16 electrodes to train the Bayesian LDA (BLDA) classifier and subsequently utilization of the trained model in an online system test. The proposed method was evaluated concerning the conventional matrix speller paradigm. The experiments were carried out on 18 subjects with ALS. Even obtaining an Acc of 90% in its most efficient performance, the results presented by the BLDA algorithm do not reveal any significant difference between the proposed approach and the conventional approach. However, patients with ALS were able to operate the system effectively.

Survival prediction studies

The survival prediction of patients with ALS is empirically defined based on, generally, the analysis of clinical data. Only one of the studies included in this review is dedicated to such a prediction. In the study developed by van der Burgh et al. [56], a model for predicting survival (short, medium, or long) of patients with ALS is proposed by combining clinical data, neuroimaging, and a robust ANN-based ML technique named Deep Learning Networks (DLN). Four scenarios were defined for the application of the DLN algorithm. Finally, the first situation was based only on clinical data. The second and third scenarios utilized MRI images and included structural connectivity and brain morphology data. The latter situation included a combination of the previous three. To each of those, a model was implemented. Furthermore, the performance of algorithms was evaluated through a database that contained data of 135 subjects with ALS. The model that combined clinical and MRI data revealed superior performance (84.4% of Acc value) and was presented as a viable strategy for predicting the survival of patients with ALS. The remaining models displayed intermediate results, although they indicated promising approaches.

Discussions

This systematic literature review explored approaches based on computational intelligence. Besides, to process biomedical signals considering the scope of ALS, it functioned in a synergistic and complementary manner. A set of 18 articles was included and reviewed, and three major classes of applications were found: aid to diagnostic, communication enabling, and survival prediction. The most adequate algorithmic models and the respective biomedical signals responsible for providing data were identified and quantified (see Fig. 6).

Fig. 6
figure 6

Quantitative of the best algorithmic models and the respective biomedical signals

Full size image

Based on the analysis of the 13 articles that addressed the support to the diagnosis of patients with ALS, regardless of the biomedical signal or ML algorithm used, it is possible to define a standard methodological scheme (a pipeline) general to all studies, which is broadly depicted in Fig. 7. Except for Khorasani et al. [49], who investigated a new classification algorithm, the studies suggest approaches or methods for the data treatment process that may enhance the training stage and, consequently, the classification stage. This data treatment process, which includes the feature extraction and selection phases, for instance, is important to eliminate noise, redundancy, and reduce the data dimensionality, in addition to maximizing the performances of the algorithms through the provision of refined and consistent data [57]. The various ML models implemented were presented as techniques for evaluating and validating the proposals of the studies. However, they were elementary techniques in the diagnosis process that are present in all articles.

Fig. 7
figure 7

Generic pipeline: generalized scheme for solving classification problems

Full size image

The studies [39, 41,42,43,44,45,46,54,55] suggest alternative approaches that include ML to optimize character selection time in a BCI system. These approaches range from the optimization of EEG electrodes to intelligent customization of the interface. The importance of BCI systems in promoting communication is evident. These systems are widely utilized in research to establish a communication pathway between the human brain and external devices, recognizing voluntary changes in the brain activity of their users [58,59,60,61,62,63].

Despite the research focused on the development of BCI systems, there are limitations regarding their home use. One of the primary reasons why BCI have not been introduced into the domestic environment is the character selection time. Specifically, the time still is considered slow and inaccurate when it comes to approaches that do not use brain signals, and also the need for electrodes connected to the head of the patient [54, 64]. Other approaches to human–computer interaction systems that do not necessarily involve brain signals by EEG can be seen in Pinheiro et al. [65], Hori et al. [66], Fathi et al. [67], Harezlak et al. [68], Villanueva et al. [69], Królak e Strumiłło [24], Zhao et al. [70], Liu et al. [71] and Aharonson et al. [22].

The only survival prediction study with ALS patients analyzes how challenging it is to develop systems for such a purpose. The study [56] indicates that MRI and the DLN technique are promising for survival prediction and suggests a more significant exploration of the field of neuroimaging. Also, the research reveals the importance and benefits of patients’ clinical data in the process of predicting survival at the three levels of ALS. This observation, in combination with the analysis made thus far, reveals both the absence and the possibility of using clinical data for diagnosis. Correlated with the survival aspect, recent studies indicate it is possible to apply ML approaches with digital biomarkers using the speech signal to monitor the progression of ALS [72], including applications for automatic classification of the ALS Functional Rating Scale (ALSFRS) [73, 74].

Regarding ML algorithms, it is observed that they are specifically supervised in all studies. The type of biomedical signal varies only in the diagnosis studies, with EMG being the most used signal, followed by GR and MRI. The EEG signal is applied solely for communication enabling applications. The MRI-based biomedical signal is used both in diagnosis and survival prediction applications. Schuster et al. [75] affirm that MRI-based biomarkers are currently seldom used for aiding the identification of ALS. This observation is complemented by the results presented in this SLR, which also reports the limited number of neuroimaging-based studies aimed at diagnosis support applications and survival prediction of the ALS disease, despite the potential mentioned by van der Burgh et al. [56]. In addition to these biomedical signs mentioned so far, studies show the feasibility of using the speech biosignal for the early diagnosis of ALS, as indicated by Wang et al. [76], Suhas et al. [77], An et al. [78], Vieira et al. [79], and Wisler et al. [80], and tracking changes in individuals with bulbar ALS [81].

The 18 studies carried out experimental tests with datasets of healthy subjects and subjects with ALS or other neurological diseases. 50% of the studies used local or proprietary datasets. The other 50% of the investigations collected data from public online repositories. In some cases, like those for diagnosis and communication, except in the study carried out by Ferraro et al. [51], the limitation in the number of patients with ALS is evident (see Fig. 4). These results suggest that it is still challenging to develop and validate a robust study with a more considerable number of subjects with ALS or in an outpatient setting.

Conclusion

This article introduces an SLR protocol to investigate relevant studies from the last ten years (2009–2019) that address ML techniques and biomedical signal processing. It may contribute to the advancement of research within the context of ALS. Based on 18 primary studies, the results exhibit strategies to minimize problems and/or promote means for diagnosis support, communication, and survival prediction. Considering the analyzed studies, 88.89% of those report the importance of treating biomedical signals for providing robust and consistent data for ML algorithmic models.

Furthermore, it can be observed that there is a predominance in the type of biomedical signals used by studies in the categories of communication and prediction of survival, being exclusively and respectively the EEG signals and MRI images. For the diagnosis class, in particular, three types of raw data are reported, namely EMG (61.54%), GR (23.08%), and MRI (15.38%). Regarding ML algorithmic models and analyzing the most satisfactory performances, SVM is the most used, followed by LDA and ANN techniques. Even though the 18 articles selected use ML, except for one study that proposed a new algorithm. In general, limited to the objectives of this SLR, the literature suggests and dedicates itself to the treatment of biomedical signals.

The studies are promising, but there are, nonetheless, significant aspects to be explored. When it comes to the diagnosis, the studies may be applied in outpatient clinics for practical assistance, in cases that have yet been unconfirmed of ALS, or even so in the early stages of the disease. Moreover, the use of big data approaches with patient’s clinical data might contribute to the conclusive results and remains open for investigations. That includes the field of survival prediction. Concerning the approaches for communication improvement, there are unanswered questions about the use of BCI in the domestic environment, considering its aspects, costs, as well as efficient interfaces that prevent fatigue, discomfort, and optimization of the electrodes for EEG signals acquisition.