Artificial Intelligence for Healthcare and Social Services: Optimizing Resources and Promoting Sustainability

Abstract

Artificial intelligence (A.I.) provides the ability to interpret massive amounts of data, which many industries are already taking advantage of. This contribution aims to investigate the potential applications of A.I. in healthcare in order to understand how it can help optimize resources in a sector that risks becoming unsustainable due to high costs and lengthy care processes. Because A.I. development is constantly evolving, the authors examined the relevant literature, focusing on the last decade to highlight the significant advances made during this time period. A scheme of uses based on the care phases is presented as a result of the analysis. This scheme, which is made up of 4 + 1 categories, can help frame and analyze potential uses. Before the conclusion, the last section of the contribution addresses the remaining challenges and discovers that there are at least three types of open issues that must be resolved before A.I. can be effectively used in healthcare, as well as other sectors. A.I may revolutionize the delivery of healthcare services, but this process must be guided because the technology does not appear to be sufficiently mature and solutions to several problems must be found.

1. Introduction

Artificial Intelligence (henceforth A.I.), among other technologies, is attracting the most interest from companies, research institutes and governments [1] because it has the potential to redefine the way multiple industries will be able to plan the services they offer, thanks to its ability to interpret huge amounts of data [2].

The term A.I. denotes the desire to create machines that can simulate the cognitive processes of the human brain, such as learning, language and problem solving [3], and use them to make choices and actions. More precisely, as defined by John McCarthy, one of the founding fathers of the discipline, A.I. is “the science and engineering of making intelligent machines” [4].

In recent years, thanks to the rapid advancement of technology, A.I. has made such strides that it has earned the limelight. A.I. has been studied since the mid-20th century, beginning with the work of Warren McCulloch and Walter Pitts [5] and especially Alan Turing [6], while the first artificial neural network is attributed to Frank Rosenblatt [7] in 1958. At the time, technology was not advanced enough to allow these theoretical studies to be put into practice [8,9]; however, as mentioned, advances in recent times, coupled with the increasing availability of data, which is essential to enable machines to learn, have given new impetus to the discipline [10].

A.I. is permeating society to such an extent that some scholars have postulated the need for a dedicated sociology of A.I. [11] to ensure that attention is paid to how values, institutional practices, and inequalities are embedded in these sociotechnical systems. In addition, this approach allows for consideration of the cultural perspective of A.I., according to which A.I. must be considered a social phenomenon that is interconnected with the cultural and economic aspects of the society in which it develops [12].

Given the potential of A.I., many sectors are looking to integrate it within their processes, for example, companies in automated warehouse management, governments using drones in military operations, but also ordinary people, eager to know what else to watch at the end of their favorite TV series.

Among the many sectors that can benefit from A.I., healthcare is one of the most promising [13] as it is usually very expensive and requires prolonged attention to users/patients. A.I., on the other hand, may be able to save resources while improving the efficiency of services [14]. The global A.I. market in healthcare is growing at a rapid pace and is expected to reach a valuation of $27.6 billion by 2025 [15].

The relationship between healthcare providers and patients is beginning to change thanks to A.I., as many services (especially information search) that were once only deliverable face-to-face no longer require human intervention and can be managed by chatbots and intelligent virtual assistants [16]. In the coming years, partly due to the increasing availability of data obtained through wearable devices [17], the use of A.I. in healthcare is expected to increase further, eventually replacing human operators in some activities [18,19]. The provision of healthcare services is likely to be revolutionized [20], but the revolution needs to be guided and accompanied because, at the moment, the technology is not sufficiently mature [21] and solutions need to be found for several problems.

This paper aims to answer the main research question, “how can A.I. be used in healthcare?” by examining the current possibilities, listing them according to stages of care, and highlighting the pitfalls yet to be addressed. Such a scheme can be really helpful in analyzing the benefits of A.I. in healthcare, as it allows researchers to focus on the different characteristics of care processes typical of each phase. It is difficult to capture all the benefits at a single glance, but if the process is broken down into different stages, this becomes easier. However, as is explained in the Results section, this subdivision into stages is somewhat incomplete because it does not take into account either the uses that span the entire care process or the uses of collateral processes. Instead, the proposed scheme, as its main strength, takes both into account.

Moreover, this study is not limited to its main purpose, as it also analyzes the main pitfalls and challenges that still need to be addressed in order to create a suitable environment for the effective implementation of A.I. in healthcare. Finally, within its methods, only recent literature was considered, so this study takes into account all relevant updates in recent years.

2. Materials and Methods

The authors drew on previous literature on the topic to catalog the uses of A.I. in healthcare and discuss the challenges that will be faced, focusing on the last decade to highlight significant advances in this time frame. This methodology was chosen because of its exploratory nature, as well as its ability to identify and synthesize evidence on a specific topic that has not been thoroughly investigated.

As a result, articles from scientific databases were chosen to answer the main research question “how can A.I. be used in healthcare?” because they are regarded as reliable analytical tools that provide quick access to relevant peer-reviewed contributions. Following the PCC (Population/Concept/Context) strategy, the authors succeeded in defining the expression of the research, using the application of Artificial Intelligence as the Concept and the various processes of healthcare services as the Context, while the Population consists of both patients and providers.

The authors conducted a search, separately, in June and July 2022, on a number of health and sociology databases: PubMED, SAGE and Springer. The search terms were conjugated using the Boolean operators “AND” and “OR” and additional tools. The final expression used was [(“artificial intelligence” OR “A.I.” OR “intelligent machines” OR “robots”) AND (“healthcare” OR “wellness” OR “health” OR “patients” OR “health practitioners”)].

The articles thus identified by each author were compared to eliminate duplicates and combined together, resulting in a corpus of more than 300 publications. Following the PRISMA method, the authors then collaborated to read the abstracts of these publications and apply inclusion and exclusion criteria to select articles for inclusion in the search. Only articles focused on the impact of A.I. on healthcare processes were included. Articles published before 2010, those not written in English, and those that were not inherent, i.e., whose study focused on A.I. or healthcare but not both, were excluded. In this way, the total number of articles considered for analysis was just over 80. From the reading of the remainder, the authors made a final selection, from which 40 studies emerged, reported in the bibliography. It is worth noting that through collaboration during the process of analyzing the studies, the authors agreed on most of the decisions regarding which studies to include in the research.

The following sections are based on the authors’ analysis of the identified studies.

3. Results: Types of A.I.

A.I. is not a single concept, but rather a set of techniques for computers to simulate human intelligence. In this way, computers can exhibit intelligent behaviors, such as learning, reasoning and problem solving.

Among the various techniques, the most important is the one called machine learning, which is the ability for computers to learn new knowledge and skills by analyzing data from the real world and how they interact with it [16]. Machine learning is currently the main approach used in A.I. and the reason why amazing results have been achieved in recent years. It is based on the development of a predictive model that can be built from already known data which then uses new data to create scenarios not considered at first, learning from one’s experience [22].

Machine learning itself is not a unique concept, as it can be divided into two types: supervised learning, which is the type described above, and unsupervised learning. The latter, which is more complex than the former, starts from unclassified (or unlabeled) data and does not rely on already known results (training sets) to fix the algorithm [23]. The success achieved by machine learning is due to its enormous potential in identifying the most effective model in each context and in processing complex data. However, to guarantee results, it needs much larger datasets. As will be discussed below, data collection is one of the potential problems of A.I., especially as it increasingly comes from wearable devices as part of the so-called Internet of Things. Smartbands and other devices are able to track biometric parameters and send them over the network to collaborate in building “big data” [24].

Another fundamental technique is Natural Language Processing (NLP), defined by Garbuio and Lin [16] as “a form of A.I. that analyzes human language, helping a machine to understand, interpret, and manipulate human language.” It analyzes human language, especially in written form, to better understand it and enable machines to fulfill requests expressed in natural language. NLP can be applied, for example, to speech recognition, machine translation, question answering, and sentiment analysis [22]. The rapid evolution of this technique, combined with speech recognition, has paved the way for the great success of intelligent virtual assistants (IVAs), led by Apple’s Siri and Amazon’s Alexa. Since voice is the most intuitive and natural way to communicate, the possibility of machines learning to use it correctly represents a breakthrough in human–computer interaction [25]. IVAs are born with this very goal in mind. At the moment, their capabilities are limited to a question/answer dialogue that must necessarily be prompted by the user; moreover, the reliability of the answers provided [26], which is crucial especially in healthcare [27], is yet to be verified, although Siri and Alexa are already able to remind people of appointments or medications to take.

Not only is A.I. composed of different techniques, but it can also have different facets, like human intelligence [28]. In their study, Pantano and Scarpi [29] identified “five main types of A.I. that show a correspondence between the human intelligences that emerged from past studies on psychology and the A.I. developed from past studies on A.I.”:

• Logical-mathematical intelligence: the ability to solve complex problems through logical thinking [30];
• Social intelligence: the ability to understand emotions and interact with human beings;
• Visual–spatial intelligence: the ability to perceive space and understand it [31];
• Linguistic–verbal intelligence: the ability to understand human language and simulate it to communicate with humans;
• Processing speed intelligence: the ability to complete simple, repetitive tasks quickly and efficiently [32,33].

Finally, A.I. can also be classified according to its level of autonomy, according to Garbuio and Lin [16]:

• Assisted intelligence is based on performing simple tasks with maximum efficiency, such as classifying diagnostic images;
• Augmented Intelligence goes beyond task repetition and modifies the very essence of a specific task to improve it and provide new skills;
• Autonomous Intelligence is the most advanced stage of A.I.; it will enable machines to make choices on their own. Substantial ethical challenges must be addressed before the actual development of this type of intelligence [34].

4. Discussion: Uses in Healthcare

Experiments are being conducted in healthcare, as in many other areas, to understand how different phases of care, from prevention to diagnosis and treatment, to rehabilitation and monitoring [35], can take advantage of the potential of A.I. The utility of these techniques in medical research and healthcare delivery is becoming clear. In The Harvard Business Review, Kalis, Collier, and Fu [36] identified several areas of healthcare that can benefit from A.I., including robot-assisted surgery, nursing, administrative workflow, dose error reduction, diagnostic accuracy improvement and radiological image interpretation. Interest in applying these techniques is also increasing financially, as investment in A.I. innovations and start-ups continues to rise [37].

As suggested at the beginning of this section, the uses of A.I. can be listed according to the stages of patient care:

• Prevention
• Diagnosis
• Treatment
• Monitoring/Rehabilitation

This list, however, is not complete; it needs another category, different from the others, because it is not specifically related to a phase of care, but to its entire process: health management, which includes, for example, how A.I. can help in the administration of a clinic. The model reported in the study by Klumpp et al. [38] also considers the category of management, called “Logistics”; however, it does not take into account all phases, including instead only two: diagnosis and treatment.

This new scheme, as its main strength, combines the advantages of what was proposed by Kassam and Kassam [35] for the division into phases, and by Reddy, Fox and Purohit [39] and Klumpp et al. [38] for the management part.

4.1. Prevention

A.I., with its unique ability to compute and interpret data, can scan huge datasets for warning signs, such as disease symptoms, proving very useful even in pandemic management [40]. By linking together data from different sources, comparing patients’ medical records with their habits and lifestyles, and cross-referencing this with historical indicators, A.I. may be able to signal in advance the risk of developing certain diseases. As reported by Fonseka, Bhat and Kennedy [41] (2019), this type of predictive technique can also be successfully applied to suicide prevention to identify people who exhibit risky behaviors.

4.2. Diagnosis

In this regard, the most important feature that A.I. can handle is clinical decision support; that is, the ability to support decisions made by physicians with clinical data analysis performed by machines. The error rate due to the human factor is greatly reduced, and consistent care processes can be ensured for patients where decisions are based on objective parameters. As a paradox, objectivity can help the personalization of care, because a choice about a specific patient can be made based on previous cases in almost identical situations [39].

4.3. Treatment

The term treatment means any type of medical intervention following the diagnosis of a disease with the goal of curing it, including surgery. As far as surgery is concerned, robotics is a particular branch of A.I. that is focusing on trying to build robot-surgeons who can operate alone or under the guidance of human operators [42].

At this stage of care, the best achievable outcome is personalization. Indeed, thanks to A.I., interventions can be tailored to the patient’s needs by studying his or her EHR, biometric data and more to provide individualized therapy while reducing the time and resources required.

4.4. Monitoring

Through monitoring, the effectiveness of what was done during treatment is verified. However, monitoring is also used to keep track of a patient’s health status remotely and, in cases of chronic diseases, can become permanent. As mentioned earlier, thanks to wearable devices and the Internet of Things, it is possible to collect a lot of data about a person’s health condition, such as sleep/wake cycles or heart rate, and the A.I. can analyze this data in real time to profile the patient and alert the, in case of problems [43]. Many devices are equipped with sensors that can detect a wide variety of parameters, thus enabling remote monitoring of patients and optimizing resources [42]. In the case of pandemics such as COVID-19, the monitoring phase is even more important: in this context, A.I. can help in operations such as contact tracing, monitoring of risky activities and more generally predicting the evolution of the disease [44]. In the monitoring phase, A.I. can be of great help. These systems can support elderly or chronic patients in their daily activities and health management, although, as mentioned, their potential is currently still limited [45].

4.5. Management

The category of management, as mentioned, is not related to a specific phase of care, like the previous ones, but rather concerns the entire process of administering healthcare services. AI can be very useful in enabling doctors, nurses, caregivers, hospitals and clinics to provide better services by reducing waste. For example, repetitive but time-consuming tasks such as patient data entry can be outsourced to A.I.-managed tools, allowing professionals to have more time to devote to patients [46]. These techniques also optimize time and resources for national health systems, improving intervention planning and the patient experience itself [42].

The usage pattern just described is not only valid in general, but can be applied in specific contexts, such as that of the COVID-19 pandemic. Indeed, as discovered by Senthilraja [44], A.I. can be used for: “(1) early warnings and alerts, (2) tracking and prediction, (3) data dashboards, (4) diagnosis and prognosis, (5) treatment and care, and (6) social control.” These areas, in fact, overlap almost completely with those described here. Fonseka, Bhat and Kennedy [41], in their study of suicides, indicate that A.I. can “improve patient care in the areas of assessment, diagnosis, treatment, and follow-up,” again using a schema that overlaps with the one used here.

5. Issues

At present, the application of A.I. to healthcare (as well as other services) is fraught with difficulties, making it less widespread than it could be [47]. To ensure that these technologies reach their full potential, at least three types of challenges must be overcome: technical, ethical, and social.

5.1. Technical Issues

The first major problem concerns the current state of the art in A.I. Despite the great achievements already made, the technology and techniques need to be further developed, as their evolution can still be considered to be in the early stages; some of the solutions adopted are not yet mature enough to be considered flawless [48]. For example, some studies have shown that some algorithms are susceptible to bias and thus may lead to erroneous results [49,50]. Technology is advancing at a rapid pace, however, and within a few years the situation could change dramatically.

The other technical problem, which is more difficult to solve, is related to the huge amount of data needed to make A.I. work. To be useful in healthcare, the data may be personal, sensitive and confidential. How is the data collected? What might happen when it is transferred? In other words, it is a matter of privacy and security. The solution to this problem, which is fundamental to the medical industry [47], is complex because it involves both consumer and professional devices and also the network infrastructure that carries the data. Furthermore, open data would be perfect for ensuring free movement [16]; obviously, however, this is impractical in light of what has just been said.

5.2. Ethical Issues

Among the ethical problems, the main one concerns the lack and fragmentation of regulation of the use of A.I. Without a clear medico–legal framework, these technologies cannot be used because liability cannot be defined [51]. If an A.I.-managed instrument makes a wrong choice that affects a patient’s health, who should be held accountable? [47].

In addition, the unreliability of current systems, already mentioned when discussing technical problems, also generates other problems related to the fairness of A.I. systems: biases in data can cause inequities based on gender, race, or other factors [52,53]. To make matters worse, data and the systems that use them can be manipulated and falsified; they can be used to falsify better results [2] and they can even for criminal purposes [54].

Finally, data obtained today can be used in the future in ways that are completely unpredictable [42].

5.3. Social Issues

Social problems are mostly related to the acceptance of A.I. by patients and professionals. The first issue in this field is trustworthiness: a device that makes data-driven decisions may seem more trustworthy; however, in reality, an ethical professional is perceived as more trustworthy for complex decisions, partly because of so-called proxy trust [55,56]. The level of trust is low because very often the algorithms that govern A.I. are not transparent, and thus are difficult for both professionals and especially patients to understand [57,58].

At a general level, the A.I. of the future is expected to be designed with more consideration of human-centered design and sustainability at the same time. For example, different designs will be necessary depending on the age of users, but also on other factors such as economic or medical conditions, in order to ensure equitable and inclusive use [59].

6. Conclusions

This study contributed to the development of a better scheme for categorizing all uses of A.I. in healthcare, based on the various stages of care, with the addition of a new category unrelated to the specific phases. The study then examined all recent updates in the literature, as well as the major pitfalls and challenges that still need to be addressed in order to create a suitable environment for the effective implementation of A.I. in healthcare.

There is no doubt that A.I. can greatly benefit healthcare management [60]. The way professionals access information and manage services has the potential to be revolutionized. Most importantly, patients can be re-centered in the process, empowered and no longer viewed solely as an “object” of care [16]. A.I., for example, can track patients’ behaviors to help them adhere to a treatment better, making them more involved in their own self-care [61].

However, A.I. should not be viewed as the panacea for all problems, as is frequently the case when a promising new technology enters the market. Indeed, as previously stated, if a thorough understanding of the process is lacking, A.I. can cause additional issues. A common concern is that A.I. will eventually take over, replacing professionals and thus eliminating the human factor, which is so important in healthcare [35]. However, for the time being, this concern is unfounded. As previously stated, Autonomous Intelligence is still a long way from becoming a reality and will not be implemented unless the challenges discussed are first addressed.

Instead, in the short to medium term, A.I. can be used exclusively under the supervision of professionals as an additional tool at their disposal, to get the best of both worlds. In this context, A.I. refers to Augmented Intelligence, or systems that aim to supplement rather than replace the methods currently used to manage healthcare at all stages [62].

Although the proposed scheme is sound and draws on previous research, the authors acknowledge that it has yet to be tested as a major limitation of the study. In future work, the authors intend to investigate this issue further and test the validity of the scheme through field research and the involvement of healthcare providers and patients.