Chatbot for Healthcare and Oncology Applications using Artificial Intelligence and Machine Learning

Background: Chatbot is a timely topic applied in various fields, including medicine and healthcare, for human-like knowledge transfer and communication. Machine learning (ML), a subset of artificial intelligence (AI), has been proven particularly applicable in healthcare with the ability for complex dialogue management and conversational flexibility. Objective: This review article reports on the recent advances and applications of chatbot technology in medicine. Methods: To provide a comprehensive background, a brief historical overview along with the developmental progress and design characteristics are first introduced. The focus will be in regards to cancer therapy with in-depth discussions and examples for diagnosis, treatment, monitoring, patient support, workflow efficiency, and health promotion. Similar with all forms of technology, risks and challenges will arise before their universal adoption in healthcare. Thus, this paper will explore limitations and areas of concern highlighting ethical, moral, security, technical, regulatory standards, and evaluation issues to explain the hesitancy of implementation. Results: Even after addressing these issues and establishing the safety or efficacy of chatbots, the human element in healthcare will not be replaceable. Therefore, chatbots have the potential to be integrated into clinical practice by working alongside health practitioners to reduce costs, refine workflow efficiencies, and improve patient outcomes. Other applications in pandemic support, global health, and education are yet to be fully explored. Conclusions: Further research and interdisciplinary collaboration could advance this technology to dramatically improve the quality of care for patients, rebalance workload for clinicians, and revolutionize the practice of medicine.


Introduction
Artificial intelligence (AI) is at the forefront of transforming numerous aspects of our lives by modifying the way we analyse information and improve decision-making through problem-solving, reasoning, and learning.Machine learning (ML) is a subset of AI that improves its performance based on data provided to a generic algorithm from experience, rather than defining rules in traditional approaches [1].Advancements in ML have provided benefits in terms of accuracy, decision making, quick processing, cost-effectiveness, and handling complex data [2].Chatbots, also known as chatter robots, smart bots, conversational agents, digital assistants, or intellectual agents, are prime examples of an AI system that has evolved from ML.The Oxford dictionary defines a chatbot as "a computer program that can hold a conversation with a person, usually over the internet".They can also be physical entities designed to socially interact with humans or other robots.Pre-determined responses are then generated by analysing user input, on text or spoken ground, and accessing relevant knowledge [3].Problems arise when dealing with more complex situations in dynamic environments and managing social conversational practices according to specific contexts and unique communication strategies [4].
Given these effectual benefits described above, it is not a surprise that chatbots have rapidly evolved these past two decades and integrated themselves into numerous fields, such as entertainment, travel, gaming, robotics, and security.Chatbots have been proven particularly applicable in various healthcare applications that usually involve face-to-face interactions.With their ability for complex dialogue management and conversational flexibility, integration of chatbot technology into clinical practice may reduce costs, refine workflow efficiencies, and improve patient outcomes [5].A Web-based, self-report survey examining physicians' perspectives found positive benefits of healthcare chatbots in managing one's own health, improved physical, psychological, and behavioural outcomes, and most notably for administrative purposes [6].In light of the opportunities provided by this relatively new technology, potential limitations and areas for concern may arise that could potentially harm users.Concerns about accuracy, cyber-security, lack of empathy, and technological maturity are reported as potential factors associated with the delay in chatbot acceptability or integration into healthcare [7].
This narrative review paper reports on the healthcare applications for chatbots with a focus on cancer therapy.The rest of the paper is organized as follows.We first introduce the developmental progress with a general overview of the architecture, design concepts, and types of chatbots.The main results section focuses on the role that chatbots play in areas related to oncology such as, diagnosis, treatment, monitoring, support, workflow efficiency, and health promotion.The discussion section analyses potential limitations and concerns for successful implementation while addressing future applications and research topics.

Methods
This review focuses on articles from peer-reviewed journals and conference proceedings.The following databases were searched from October to December 2020 for relevant and current studies from 2000 to 2020: IEEE Xplore, PubMed, Web of Science, Scopus, and OVID.The literature search used the following key terms ('chatbot', 'chatter robot', 'conversational agent', 'artificial intelligence', 'machine learning').For further refinement, these key terms were combined with more specific terms aligned with the focus of the article.This included 'healthcare', 'cancer therapy', 'oncology, 'diagnosis', 'treatment', 'radiation therapy', 'radiotherapy'.The searches were not limited by language or study design.Letters and technical reports were excluded from the search.The full list of sources and search strategy is available from the authors.
Screening of chatbots was guided by a systematic review process from the Botlist directory during the period of January 2021.This directory was chosen because it was openaccess, categorized the chatbots under many different categories (ie.healthcare, communication, entertainment, etc.), and contained many commonly used messenging services (ie.Facebook Messenger, Discord, Slack, Kik, Skype).A total of 78 chatbots were identified for healthcare application and further divided according to following criteria: diagnosis, treatment, monitoring, support, workflow, health promotion.It should be noted that using the health filters from a web directory limits the results to searching strategy and marketing label.Thus, results from equivalent studies may differ when repeated.

Chatbot History and Evolution
The idea of a chatbot was first introduced in 1950 when Alan Turing proposed the question, "Can machines think?" [8].The earliest forms were designed to pass the Turing test and mimic human conversations as much as possible.In 1966, Eliza was the first known chatbot developed that acted as a psychotherapist using pattern matching and template-based responses to converse in a question format [9].Improvements were made to build a more human-like and personalized entity by incorporating a personality in PARRY that simulated a paranoid patient [10].One of the most well-known chatbot is ALICE, developed in 1995, that uses a pattern-matching technique to retrieve example sentences from output templates and avoid inappropriate responses [11].A renewed interest in artificial intelligence and advances in machine learning have led to the growing use and availability of chatbots in various fields [12].SmarterChild [13] became widely accessible through messenger applications followed by more familiar virtual assistants using voice-activated systems, such as Apple Siri, Amazon Alexa, Google Assistant, or Microsoft Cortana.Based on our analysis (Figure 1), the most popular development of chatbots for healthcare purposes are diagnostics, patient support (ie. mental health counselling), and health promotion.Some of these applications will be further explored in the following section for cancer applications.

Chatbot General Architecture
Although there are a variety of techniques for the development of chatbots, the general layout is relatively straightforward.As a computer application that uses machine learning to mimic human conversation, the underlying concept is similar between all types with four essential stages (input processing, input understanding, response generation, and response selection) [14].A simplified general chatbot architecture is illustrated below (Figure 2).First, the user makes a request, in text or speech format, that is received and interpreted by the chatbot.From there, the processed information could be remembered or more details could be requested for clarification.After the request is understood, the requested actions are performed and the data of interest is retrieved from the database or external sources [15].

Chatbot Types
With the vast amount of algorithms, tools, and platforms available, understanding the different types and end purpose of these chatbots will assist developers when choosing the optimal tools when designing them to fit the specific needs of users.These categories are not exclusive as chatbots may possess multiple characteristics making the process more variable.The five main types are described below [15].Table 1 describes some examples of the recommended applications for each type of chatbot, but are not limited to the ones specified.
Knowledge domain classification is based on accessible knowledge or the data used to train the chatbot.Under this category are the open domain for general topics and the closed domain focusing on more specific information.Service provided classification is dependent on the sentimental proximity to the user and the amount of intimate interaction dependent on the task performed.This can be further divided into interpersonal providing services to transmit information, intrapersonal for companionship or personal support to humans, and inter-agent to communicate with other chatbots [14].The next classification is based on goals with the aim to achieve, sub-divided to informative, conversational, and task-based.Response generation chatbots, further classified as rule-based, retrieval-based, and generative, account for the process of analyzing inputs and generating responses [16].Lastly, the human-aided classification incorporates human computation which provides more flexibility and robustness, but lacks the speed to accommodate more requests [17].Table 1.Recommended healthcare applications for the different types of chatbots.

Knowledge Domain
Open: Responding to more general and broader topics that can be easily searched within databases.May be the preferred chatbot type for routine symptom screening, connecting to providers/services, or health promotion applications.
Closed: Responding to complex or specific questions requiring more in-depth research.May be the preferred chatbot type for treatment planning or recommendation.

Service Provided
Interpersonal: Used mainly to transmit information without much intimate connection with users.May be the preferred chatbot type for imaging diagnostics or hereditary assessment where the main duty is to relay factual information to users.
Intrapersonal: Tailored for companionship or support.May be the preferred chatbot type for counselling, emotional support, or health promotion that requires a sense of human touch.
Inter-agent: Used for communicating with other chatbots or computer systems.May be the preferred chatbot type for administration purposes when transferring patient information between locations.

Goal-based
Informative: Designed to provide information from warehouse database or inventory entry.May be the preferred chatbot type for connecting patients with resources or remote patient monitoring.
Conversational: Built with the purpose of conversing with users as naturally as possible.May be the preferred chatbot type for counselling, emotional support, or health promotion.
Task-based: Only performs one specific task where actions are predetermined.May be the preferred chatbot type for screening and diagnostics.

Response Generation
Uses pattern matching when the domain is narrow and sufficient data is available to train the system.May be the preferred chatbot type for screening and diagnostics.

Human Aided
Incorporates human computation that increases flexibility and robustness, but decreases speed.May be the preferred chatbot type for most applications, except for support or workflow efficiency where speed is an essential factor in the delivery of care.

Chatbots in Cancer Therapy
Cancer has become a major health crisis and is the second leading cause of death in the US [18].The exponentially increasing number of cancer patients each year may be a combination of carcinogens in the environment and improved quality of care.The latter aspect could explain why cancer is slowly becoming a chronic disease that is manageable over time [19].The added life expectancy poses new challenges for both patients and the healthcare team.For example, many patients now require extended at-home support and monitoring while healthcare workers deal with the increased workload.Although clinician's knowledge-base has exploded in the use of scientific evidence to guide decision making, there are still many other facets to the quality of care that has yet to catch up.Key areas of focus are the safety, effectiveness, timeliness, efficiency, equitability, and patient-centered care [20].
Chatbots have the potential to address many of the current concerns of cancer care mentioned above.This includes the triple aim of healthcare that encompasses improving the experience of care, improving the health of populations, and reducing per capita costs [21].Chatbots can improve the quality or experience of care by providing efficient, equitable, and personalized medical services.We can think of them as intermediaries with physicians to facilitate history taking of sensitive and intimate information before consultations.They could also be thought of as decision aids that deliver regular feedback on disease progression and treatment reaction to help clinicians better understand individual conditions.Preventative measures of cancer have become a global priority because early detection and treatment alone have not been effective in eliminating this disease [22].Physical, psychological, and behaviour improvements of underserved or vulnerable populations may even be possible through chatbots because they are so readily accessible through common messaging platforms.Health promotion usage, such as lifestyle coaching, healthy eating, and smoking cessation, has been one of the most common chatbots according to our search.Additionally, chatbots could help save a significant amount of healthcare costs and resources.Newer therapeutic innovations have come with a heavy price tag and out-of-pocket expenses have placed a significant strain on patients' financial well-being [23].With chatbots implemented into cancer care, consultations for minor health concerns may be avoided which allows clinicians to spend more time with patients who need their attention the most.Costs may also be reduced by delivering medical services more efficiently.For example, the workflow could be streamlined by assisting physicians in administrative tasks, such as scheduling appointments, providing medical information, or locating clinics.
With the rapidly increasing applications of chatbots to healthcare, this section will explore several areas of development and innovation in cancer care.Various examples of current chatbots provided below will illustrate their ability to tackle the triple aim of healthcare.The specific use case of chatbots in oncology with examples of actual products and proposed designs are outlined below (Table 1).Examines radiological images to aid clinicians with diagnosis Symptom screening Quro [25] Pre-synopsis based on symptoms and history to predict user conditions Buoy Health [26] Assists in identifying the cause of illnesses and provides medical advice Harshitha breast cancer screening [27] Dialog flow to give an initial analysis of breast cancer symptoms Babylon [28] Symptom checker Your.md [28] Symptom checker Ada [28] Symptom checker

Hereditary assessment
ItRuns [29] Gathers family history information at the population level to determine the risk of hereditary cancer Treatmen t Patient treatment recommendatio n Mathew [30] Identifies symptoms, predicts the disease using a symptom-disease dataset, and recommends a suitable treatment Madhu [31] Provides list of available treatments for various diseases, informs the user of the composition and prescribed use of the medications Connecting patients with providers/resour ces Divya [32] Engages patients about their symptoms to provide a personalized diagnosis and connects with appropriate medical service Rarhi [33] Provides a diagnosis based on symptoms, measure the seriousness, and connect with a physician Physician treatment planning

Watson
for Oncology [34] Examines data from records and medical notes to generate an evidence-based treatment plan for oncologists Monitorin g

Remote patient monitoring
STREAMD [35] Provide access to care instructions and educational information Conversa [35] Provide access to care instructions and educational information Memora Health [35] Provide access to care instructions and educational information AiCure [36] Coach patients in managing their condition and adhering to instructions Infinity [37] Assess health outcomes and impact of phone-based monitoring for cancer patients 65 + years Vik [38,39] Addresses patient's daily needs and concerns Support Counselling Vivobot [40] Cognitive and behavioural intervention for positive psychology skills and promote well-being Emotional support Youper [26] Daily emotional support and mental health tracking Wysa [26] Daily emotional support and mental health tracking Replika [26] Daily emotional support and mental health tracking Unmind [26] Daily emotional support and mental health tracking Shim [26] Daily emotional support and mental health tracking Woebot [41] Daily emotional support and mental health tracking Workflow Efficiency

Administration
Sense.ly [42] Assist in monitoring appointments, manage patients' conditions, and suggest therapies Careskore [42] Tracks vitals and anticipates the need for hospital admissions Mandy [43] Assists healthcare staff by automating the patient intake process Patient encounter HOLMeS [44] Support diagnosis, choose the proper treatment pathway, and provide prevention check-ups Health Promotio n

General lifestyle coaching
SWITCHes [45] Track patients' progress, provides insight to physicians, and suggests suitable activities CoachAI [46] Track patients' progress, provides insight to physicians, and suggests suitable activities WeightMentor [47] Provides self-help motivation for weight loss maintenance and allows for open conversation Healthy eating Health Hero [48] Guides in making informed decisions around food choices to change unhealthy eating habits Tasteful Bot [48] Guides in making informed decisions around food choices to change unhealthy eating habits Forksy [48] Guides in making informed decisions around food choices to change unhealthy eating habits SLOWbot [49] Guides in making informed decisions around food choices to change unhealthy eating habits Smoking cessation SMAG [50] Cognitive behaviour therapy Bella [51] Coach to help quit smoking

Diagnostics and Screening
Receiving an accurate diagnosis is critical for appropriate care to be administered.In terms of cancer diagnostics, AI-based computer vision is a function often used in chatbots that could recognize subtle patterns from images.This would increase physicians' confidence when identifying cancer types because even highly trained individuals may not always agree on the diagnosis [52].Studies have shown that the interpretation of medical images for the diagnosis of tumours performs equally as well or better with AI compared to experts [53][54][55][56].Additionally, automated diagnosis may be useful when there are not enough specialists to review images.This was made possible through deep learning algorithms in combination with the increasing availability of databases for the tasks of detection, segmentation, and classification [57].For example, Medical Sieve is a chatbot that examines radiological images to aid and communicate with cardiologists and radiologists to identify issues quickly and reliably [24].Similarly, Microsoft's InnerEye is a computer-assisted image diagnostic chatbot that recognizes cancers and diseases within the eye, but does not directly interact with the user like a chatbot [42].Even with the rapid advancements of AI in cancer imaging, a major issue is the lack of a gold standard [58].
From the patient's perspective, various chatbots have been designed for symptom screening and self-diagnosis.The ability for patients to be directed to urgent referral pathways through early warning signs has been a promising market.Decreased wait times in accessing healthcare services have been found to correlate with improved patient outcomes and satisfaction [59][60][61].The automated chatbot, Quro, provides pre-synopsis based on symptoms and history to predict user conditions (average precision ~ 0.82) without a form-based data entry system [25].In addition to diagnosis, Buoy Health assists users in identifying the cause of their illness and provides medical advice [26].Another chatbot designed by Harshitha et al. uses dialog flow to give an initial analysis of breast cancer symptoms.It has been proved to be 95% accurate in differentiating between normal and cancerous images [27].Even with the promising results, there are still potential areas for improvement.A study of three mobile-app based chatbot symptom checkers (Babylon, Your.me, and Ada) indicated that sensitivity remained low at 33% for the detection of head and neck cancer [28].The number of studies assessing the development, implementation, and effectiveness are still relatively limited compared to the diversity of chatbots currently available.More studies are required to establish efficacy across various conditions and populations.Nonetheless, chatbots for selfdiagnosis is an effective way to advise patients as the first point of contact if accuracy and sensitivity requirements can be satisfied.
Early cancer detection can lead to higher survival rates and improved quality of life.Inherited factors are present in 5-10% of cancers, including breast, colorectal, prostate, and rare tumour syndromes [62].Family history collection is a proven way to easily access the genetic disposition of developing cancer to inform risk-stratified decision making, clinical decisions, and cancer prevention [63].Web-based chatbot, ItRuns, gathers family history information at the population level to determine the risk of hereditary cancer [29].We have yet to find a chatbot that incorporates deep learning to process large and complex datasets at a cellular level.Although not able to directly converse with users, DeepTarget [64] and deepMirGene [65] are capable of performing miRNA and target prediction using expression data with higher accuracy compared to non-deep learning models.With the advent of phenotype-genotype predictions, chatbots for genetic screening would greatly benefit from image recognition.New screening biomarkers are also being discovered at a rapid speed, so continual integration and algorithm training are required.These findings align with studies that demonstrate chatbots have the potential to improve user experience, accessibility, and provide accurate data collection [66].

Treatment
Chatbots have now been able to provide patients with treatment and medication information after diagnosis, without having to directly contact a physician.Such a system was proposed by Mathew et al. that identifies the symptoms, predicts the disease using a symptomdisease dataset, and recommends a suitable treatment [30].Although this may seem like an attractive option for patients looking for a fast solution, computers are still prone to errors and bypassing professional inspection may be an area of concern.Chatbots may also be an effective resource for patients who want to learn why a certain treatment is necessary.Madhu et al. proposed an interactive chatbot application that provides a list of available treatments for various diseases, including cancer.This system also informs the user of the composition and prescribed use of the medications to help select the best course of action [31].The diagnosis and course of treatment for cancer are complex, so a more realistic system would be a chatbot used to connect users with the appropriate specialists or resources.A text-to-text chatbot by Divya et al. engages patients about their medical symptoms to provide a personalized diagnosis and connects the user with the appropriate physician if major diseases are detected [32].Rarhi et al. proposed a similar design that provides a diagnosis based on symptoms, measure the seriousness, and connect users with a physician if needed [33].In general, these systems may greatly help individuals in conducting daily check-ups, increase awareness of their health status, and encourage users to seek medical assistance for early intervention.
Chatbots have also been used by physicians during treatment planning.An example is IBM's Watson for Oncology that examines data from records and medical notes to generate an evidence-based treatment plan for oncologists [34].Studies have shown that Watson for Oncology still cannot replace experts at this moment because quite a few cases are not consistent with experts (~73% concordant) [67,68].Nonetheless, this could be an effective decision-making tool for cancer therapy to standardize treatments.Although not specifically an oncology application, another chatbot example for clinician's use is the chatbot Safedrugbot [69].This is a chat messaging service for health professionals offering assistance about appropriate drug use information during breastfeeding.Promising progress has also been made for using AI in radiotherapy to reduce the workload of radiation staff or identify at-risk patients by collecting outcomes before and after treatment [70].An ideal chatbot for healthcare professionals' use would be able to accurately detect diseases and provide the proper course of recommendations, which are functions currently limited by time and budgetary constraints.Continual algorithm training and updates would be necessary because of the constant improvements with current standards of care.Further refinements and testing for accuracy to algorithms are required before clinical implementation [71].This area holds tremendous potential as an estimated 50% or more of all cancer patients have used radiotherapy during the course of their treatment.

Patient Monitoring
Chatbots have been implemented in remote patient monitoring for post-operative care and follow-ups.The healthcare sector is among the most overwhelmed by those needing continued support outside hospital settings as the majority of patients newly diagnosed with cancer are 65 years or older [72].The integration of this application would improve patients' quality of life and relieve the burden on healthcare providers through better disease management, reducing the cost of visits, and allowing timely follow-ups.In terms of cancer therapy, remote monitoring can support patients by enabling higher dose chemotherapy drug delivery, reducing secondary hospitalizations, and providing health benefits after surgery [73][74][75].
STREAMD, Conversa, and Memora Health are chatbots that function on existing messaging platforms that provide patients with immediate access to care instructions and educational information [35].To ensure patients are adhering to instructions, AiCure uses a smartphone webcam to coach them in managing their condition.Recently, a chatbot architecture was proposed for patient support based on microservices to provide personalized eHealth functionalities and data storage [36].Several studies have supported the application of chatbots for patient monitoring [76].The semi-automized messaging chatbot, Infinity, was used to assess the health outcomes and healthcare impact of phone-based monitoring for cancer patients 65 years and older.After two years of implementation, there was a 97% satisfactory rate and 87% considered monitoring useful with the most reported benefit due to treatment management and moral support [37].Similar results were discovered from two studies using Vik, a text-based chatbot that responds to the daily needs and concerns of patients and their relatives with personal insights.A one-year prospective study of 4737 breast cancer patients had a 94% overall satisfaction rate [38].More in-depth analysis of the 132,970 messages showed that users were more likely to answer multiple choice questions compared to openended ones, chatbots improved treatment compliance rate by over 20% (p=0.04), and intimate or sensitive topics were openly discussed.An area of concern is that retention rates drastically decreased to 31% by the end of this study.The other study was a phase three, blind, noninferiority randomized controlled trial (n=132) to assess the level of patient satisfaction of answers provided by chatbots vs. physicians [39].Using 12 frequently asked questions about breast cancer, participants were split into two groups to rate the quality of answers from chatbots or physicians.Among breast cancer patients in treatment or remission, chatbots answers were shown to be non-inferior (p<0.001) with a success rate of 69% compared to 64% in physician groups.Concerns about the chatbot's ability to successfully answer more complex questions or detecting differences between major and minor symptoms still remain to be addressed.
Further refinements and large scale implementation is still required to determine the benefits across different populations and sectors in healthcare [26].Although overall satisfaction is found to be relatively high, there is still room for improvement by taking into account user feedback tailored to the patient's changing needs during recovery.In combination with wearable technology and affordable software, chatbots have great potential to impact patient monitoring solutions.

Patient Support
The prevalence of cancer is increasing along with the number of cancer survivors partly due to the improved treatment techniques and early detection [77].These individuals experience added health problems, such as infections, chronic diseases, psychological issues, and sleep disturbances, which often require specific needs not able to be met by many practitioners (ie.medical, psychosocial, informational, and proactive contact) [78].A number of these individuals require support after hospitalization or treatment periods.Maintaining autonomy and living in a self-sustaining way within their home environment is especially important for older populations [79].Implementation of chatbots may address some of these concerns, such as reducing the burden on the healthcare system and supporting independent living.
With psychiatric disorders affecting at least 35% of cancer patients, comprehensive cancer care now includes psychosocial support to reduce distress and foster a better quality of life [80].The first chatbot was designed for individuals with psychological issues [9], but they continue to be used for emotional support and psychiatric counselling with their ability to express sympathy and empathy [81].Health-based chatbots delivered through mobile applications, such as WoeBot, Youper, Wysa, Replika, Unmind, and Shim offer daily emotional support and mental health tracking [26].A study performed on Woebot, developed based on cognitive behavioural therapy, showed that depressive symptoms were significantly reduced and participants were more receptive than traditional therapies [41].This agreed with the Shim results, also using the same type of therapy, which showed that the intervention was highly engaging, improved well-being, and reduced stress [82].When another chatbot was developed based on the structured association technique counselling method, the user's motivation was enhanced and stress was reduced [83].Similarly, a graph-based chatbot has been proposed to identify the mood of users through sentimental analysis and provide humanlike responses to comfort patients [84].Vivobot provides cognitive and behavioural intervention to deliver positive psychology skills and promote well-being.This psychiatric counselling chatbot was effective in engaging users and reduced anxiety for young adults after cancer treatment [40].The limitation to the studies above was that the majority of participants were young adults, most likely due to the platform the chatbots were available on.Additionally, longer follow-up periods with larger and more diverse sample sizes are needed for future studies.Chatbots used for psychological support hold great potential because individuals are more comfortable disclosing personal information when no judgements are formed, even if users could still discriminate their responses from that of humans' [82,85].

Workflow Efficiency
Electronic health records have improved data availability, but also increased the complexity of clinical workflow contributing to ineffective treatment plans and uninformed management [86].A streamlined process using ML techniques would allow clinicians to spend more time with patients by decreasing the time spent on data entry through the ease of documentation, exposing relevant patient information from the chart, automatically authorizing payment, or reducing medical errors [58].For example, Mandy is a chatbot that assists healthcare staff by automating the patient intake process.Using a combination of datadriven natural language processing with knowledge-driven diagnostics, this chatbot interviews the patient, understands their chief complaints, and submits reports to physicians for further analysis [43].Similarly, Sense.lyacts as a virtual nurse to assist in monitoring appointments, manage patients' conditions, and suggest therapies.Another chatbot that reduces the burden on clinicians and decreases wait times is Careskore that tracks vitals and anticipates the need for hospital admissions [42].Chatbots have also been proposed to autonomize the patient encounters through several advanced eHealth services.In addition to collecting data and providing bookings, Health On-Line Medical Suggestions (HOLMeS) interacts with the patients to support diagnosis, choose the proper treatment pathway, and provide prevention check-ups [44].Although the use of chatbots in healthcare and cancer therapy has the potential to enhance clinician efficiency, reimbursement codes for practitioners are still lacking before universal implementation.Additionally, studies will be to be conducted to validate the effectiveness of chatbots in streamlining workflow for different healthcare settings.Nonetheless, chatbots hold great potential to complement telemedicine by streamlining medical administration and autonomizing patient encounters.

Health Promotion
Cancer survivors, particularly those who underwent treatment during childhood, are more susceptible to adverse health risks and medical complications.Consequently, promoting a healthy lifestyle early on is imperative to maintain quality of life, reduce mortality, and decrease the risk of secondary cancers [87].According to the analysis from the web directory, heath promotion chatbots are the most commonly available, but most of them are only available on a single platform.Thus, interoperability on multiple common platforms is essential for adoption by various types of users across different age groups.Additionally, voice and image recognition should also be considered as most are still text-based.
Healthy diets and weight control are key to successful disease management as obesity playing a significant risk for chronic conditions.Chatbots have been incorporated into health coaching systems to address health behaviour modification.For example, CoachAI and Smart Wireless Interactive Health System (SWITCHeS) track patients' progress, provides insight to physicians, and suggest suitable activities [45,46].Another application is Weight Mentor that provides self-help motivation for weight loss maintenance and allows for open conversation without being affected by emotions [47].Health Hero, Tasteful Bot, Forksy, and SLOWbot guides users to make informed decisions around food choices to change unhealthy eating habits [48,49].The effectiveness of these applications can not be concluded because more rigorous analysis of the development, evaluation, and implementation is required.Nevertheless, chatbots are emerging as a solution for healthy lifestyle promotion through access and human-like communication while maintaining anonymity.
Most would assume cancer survivors would be more inclined to practice health protection behaviours with the extra guidance from health professionals, but results have been surprising.Smoking accounts for at least 30% of all cancer deaths, but up to 50% of survivors continue to smoke [88].The benefit of using chatbots for smoking cessation across various age groups has been highlighted in numerous studies showing improved motivation, accessibility, and adherence to treatment which have led to increased smoking abstinence [89][90][91].Cognitive behaviour therapy-based chatbot, SMAG, supporting users over Facebook social network resulted in 10% high cessation rate compared to control groups [50].Motivational interviewbased chatbots have been proposed with promising results, where a significant number of subjects showed an increase in their confidence and readiness to quit smoking after one week [92].No studies have been found to access the effectiveness of chatbots for smoking cessation in terms of ethnic, racial, geographic, or socioeconomic status differences.Creating chatbots with pre-specified answers is simple, but the problem becomes more complex when answers are open.Bella, one of the most advanced text-based chatbots on the market advertised as a coach for adults, gets stuck when responses are not prompted [51].Therefore, reacting to unexpected responses is still an area in progress.Given all the uncertainties, chatbots hold potential for those looking to quit smoking because they prove to be more acceptable for users when dealing with stigmatized health issues compared to general practitioners [7].

Challenges and Limitations
Advances in AI and ML have moved forward at impressive rates and revealed the potential of chatbots in healthcare and clinical settings.AI technology outcompetes humans in terms of image recognition, risk stratification, improved processing, and 24/7 assistance with data and analysis.However, there is no machine substitute for higher-level interaction, critical thinking, and ambiguity [93].Chatbots create added complexity that must be identified, addressed, and mitigated before their universal adoption in healthcare.
Hesitancy from physicians and poor adoption by patients is a major barrier to overcome, which could be explained by many of the factors discussed in this section.A cross-sectional web-based survey of 100 practicing physicians gathered the perceptions of chatbots in healthcare [6].Although a wide variety of beneficial aspects were reported (i.e.management of health and administration), an equal amount of concerns were present.Over 70% of physicians believed that chatbots cannot effectively care for all of the patients' needs, cannot display human emotion, cannot provide detailed treatment plans, and poses a risk if patients selfdiagnose or do not fully comprehend their diagnosis.If the limitations of chatbots are better understood and mitigated, the fears of adopting this technology into healthcare may slowly subside.We end the discussion by exploring challenges and questions for healthcare professionals, patients, and policymakers.

Moral and Ethical Constraints
The use of chatbots in healthcare presents a novel set of moral and ethical challenges that must be addressed for the public to fully embrace this technology.Issues to consider are privacy or confidentiality, informed consent, and fairness.Each one of these concerns will be addressed below.Although efforts have been made to address these concerns, current guidelines and policies are still far behind the rapid technological advances [94].
Healthcare data is highly sensitive due to the risk of stigmatization and discrimination if the information is wrongfully disclosed.The ability for chatbots to ensure privacy is especially important as vast amounts of personal and medical information are often collected without users being aware, including voice recognition and geographical tracking.The public's lack of confidence is not surprising given the increased frequency and magnitude of high-profile security breaches and inappropriate use of data [95].Unlike financial data that becomes obsolete after being stolen, medical data are particularly valuable because they are not perishable.Privacy threats may break the trust that is essential to the therapeutic physicianpatient relationship and inhibits open communication of relevant clinical information for proper diagnosis and treatment [96].
Chatbots experience the "Black Box" problem, similar to many computing systems programmed using ML that are trained on massive data sets to produce multiple layers of connections.Although they are capable of solving complex problems unimaginable by humans, these systems remain highly opaque and the resulting solutions may be unintuitive.This means that the systems' behaviour is hard to explain by merely looking inside and understanding exactly how they are programmed is near impossible.For both users and developers, transparency becomes an issue because they aren't able to fully understand the solution or intervene to predictably change the chatbot's behaviour [97].With the novelty and complexity of chatbots, obtaining valid informed consent where patients can make their own healthrelated risk and benefit assessments becomes problematic [98].Without sufficient transparency, deciding how certain decisions were made or how errors may occur reduces the reliability of the diagnostic process.The "Black Box" problem also poses a concern to patient autonomy by potentially undermining the shared decision-making between physicians and patients [99].The chatbot's personalized suggestions are based on algorithms and refined based on the user's past responses.The removal of options may slowly reduce the patient's awareness of alternatives and interfere with free choice [100].
Lastly, the issue of fairness arises with algorithm bias when data used to train and test chatbots do not accurately reflect the people they represent [101].As the AI field lacks diversity, the bias at the level of the algorithm and modelling choices may be overlooked by developers [102].In a study using two cases, differences in prediction accuracy were shown concerning gender and insurance type for ICU mortality and psychiatric readmissions [103].On a larger scale, this may exacerbate barriers to healthcare for minorities or underprivileged individuals leading to worse health outcomes.Identifying the source of algorithm bias is crucial to address healthcare disparities between various demographic groups and improve data collection.

Chances for Errors
Although studies have shown that AI technologies make fewer mistakes compared to humans in terms of diagnosis and decision making, they still bear inherent risks for medical errors [104].The interpretation of speech remains prone to errors due to the complexity of background information, accuracy of linguistic unit segmentation, variability in acoustic channels, and linguistic ambiguity with homophones or semantic expressions.Chatbots are unable to efficiently cope with these errors because of the lack of common sense and inability to properly model real-world knowledge [105].Another factor that contributes to errors and inaccurate predictions is the large noisy data sets used to train modern models because large quantities of high quality, representative data are often unavailable [58].Aside from the concern of accuracy and validity, addressing clinical utility and effectiveness of improving patients' quality of life is just as important.With the increased use of diagnostic chatbots, the risk of over-confidence and overtreatment may cause more harm than benefit [99].There still holds clear potential for improved decision-making as diagnostic deep learning algorithms were found to be equivalent to healthcare professionals in classifying diseases in terms of accuracy [106].These issues presented above all raise the question of who is legally liable for medical errors.Avoiding responsibility becomes easier when numerous individuals are involved at multiple stages from development to clinical applications [107].Although the law has been lagging and litigation is still a grey area, determining legal liability becomes increasingly pressing with chatbots becoming more accessible in healthcare.

Regulatory Considerations
Regulatory standards have been developed to accommodate for the rapid modifications along with ensuring the safety and effectiveness of AI technology, including chatbots.The US Food and Drug Administration (FDA) has recognized the distinctiveness of chatbots compared to traditional medical devices by defining the software within the medical device category and outlined its approach through the Digital Health Innovation Action Plan [108].With the growing number of AI algorithms approved by FDA, they opened public consultation for setting performance targets, monitoring performance, and reviewing when performance strays from pre-set parameters [102].American Medical Association (AMA) has also adopted the Augmented Intelligence in Health Care policy for the appropriate integration of AI into healthcare by emphasizing the design approach and enhancement of human intelligence [109].One area of concern is that chatbots are not covered under the Health Insurance Portability and Accountability Act (HIPAA), so users' data may be unknowingly sold, traded, and marketed by companies [110].On the other hand, over-regulation may diminish the value of chatbots and decrease the freedom for innovators.Consequently, balancing these opposing aspects is essential to promote the benefits and reduce the harms to the healthcare system and society.

Future Directions
Chatbots' robustness of integrating and learning from large clinical data sets along with its ability to seamlessly communicate with users contributes to its widespread integration in various healthcare applications.Given the current status and challenges of cancer care, chatbots will likely be a key player in this field's continual improvement.More specifically, they hold promise in addressing the triple aim of healthcare by improving the quality of care, bettering the health of populations, and reducing the burden or cost of our healthcare system.Beyond cancer care, there are increasing numbers of creative ways chatbots could be applicable in healthcare.During the COVID-19 pandemic, chatbots have already been deployed to share information, suggest behaviour, and offer emotional support.They have the potential to prevent misinformation, detect symptoms, and lessen the mental health burden during global pandemics [111].On a global health level, chatbots have emerged as a socially responsible technology to provide equal access to quality healthcare and break down the barriers between the rich and poor [112].To further advance medicine and knowledge, the use of chatbots in education for learning and assessments is crucial to providing objective feedback, personalized content, and cost-effective evaluations [113].For example, the development of the Einstein application as a virtual physics teacher enables interactive learning and evaluations, but is still far from being perfect [114].Given chatbots' diverse applications in numerous aspects of healthcare, further research and interdisciplinary collaboration to advance this technology could revolutionize the practice of medicine.
Based on the discussion above, the following features are general directions of future suggestions for improvements in chatbots within cancer care in no particular order of importance:  Cancer patients may feel vulnerable or fear discrimination from employers or society [115].Security of sensitive information must be held to the highest standards, especially when personal health information is shared between providers and hospital systems.
 An increasing number of patients are bringing internet-based information to consultations that are not critically assessed for trustworthiness or credibility.If used correctly, the additional health information could enhance understanding, improve the ability to manage their conditions, and increase confidence during interaction with physicians [116].Unfortunately, this is often not the case and most patients are not adequately informed about the proper screening of information.Ways to address this challenge is to promote awareness and develop patient management guidelines.
Chatbots also have the potential to become a key player in their ability to screen for credible information.They could help vulnerable individuals critically navigate online cancer information, especially for the older or more chronic populations that tend to be less technologically adept.
 Current applications of chatbots as computerized decision support systems in diagnosis and treatment are relatively limited.The targeted audience for most have been for patients' use and few are designed to aid physicians at the point of care.Medical Sieve and Watson for Oncology are the only chatbots found in our search designed specifically for clinicians.There are far more AI tools on the market to help with clinical decisionmaking without the ability to interact with users [117].With the rapid data collection from electronic health records, real-time predictions, and links to clinical recommendations, adding chatbot functionalities to current decision aids will only improve upon patient-centered care and streamline workflow for clinicians.
 More concrete evidence of high quality and accuracy across a broad range of conditions and populations.This entails more representative training data reflecting racial biases and developing peer-reviewed algorithms to reduce the "Black Box" problem.
 Integration into healthcare system, particularly with telemedicine, for seamless delivery from beginning to end.This does not mean replacing in-person care, but rather complementing healthcare workflow to ensure patients receive continuity and coordination of care.
 Reimbursement of chatbot services to physicians who decide to implement this technology to their practice will likely increase adoption rates.Organizations and health providers will likely profit since chatbots allow for a more efficient and reduce cost of delivery.
 Continual training of chatbots as new knowledge is uncovered, such as symptom patterns or standard of care.
 Since the Vik study found that users were more likely to respond to multiple choice questions over open-ended ones [38], chatbot developers should move towards the choice with higher response rates.Studies, survey, and focus groups should continue to be conducted to determine the best ways to converse with users.
 Universal adoption of various technical features: training with additional languages, image recognition, voice recognition, user feedback to improve services according to needs, access on multiple common platforms, and reacting to unexpected responses.
The ability to accurately measure performance is critical for the continual feedback and improvement of chatbots, especially the high standards and vulnerable individuals served in healthcare.Given that the introduction of chatbots to cancer care is relatively recent, rigorous evidence-based research is lacking.Standardized indicators of success between users and chatbots need to be implemented by regulatory agencies before adoption.Once the primary purpose is defined, common quality indicators to consider are the success rate of a given action, non-response rate, comprehension quality, response accuracy, retention/adoption rates, engagement, and satisfaction level.The ultimate goal is to assess whether chatbots positively impacted and addresses the three aims of healthcare.Regular quality checks are especially critical for chatbots acting as a decision aid because they could have a major impact on patients' health outcomes.

Review Limitations
The systematic literature review and chatbot database search includes a few limitations.The literature review and chatbot search were all conducted by a single reviewer, which could potentially have introduced bias and limited findings.Additionally, our review explored a broad range of healthcare topics and some areas could have been elaborated upon and explored more deeply.Furthermore, only a limited number of studies were included for each subtopic of chatbot for oncology applications due to the scarcity of studies addressing this topic.Future studies should consider refining the search strategy to identify other potentially relevant sources that may have been overlooked and assign multiple reviews to limit individual bias.

Conclusion
As illustrated in this review, these chatbots' potential in cancer diagnostics and treatment, patient monitoring and support, clinical workflow efficiency, and health promotion have yet to be fully explored.Numerous risks and challenges will continue to arise that require careful navigation with the rapid advancements of chatbots.Consequently, weighing the gains versus threats with a critical eye is imperative.Even after laying down the proper foundations for using chatbots safely and effectively, the human element in the practice of medicine is irreplaceable and will always be present.Healthcare professionals have the responsibility to understand both the benefits and risks associated with chatbots, and in turn educate their patients.

Figure 1 .
Figure 1.Search and screening for healthcare chatbots.Applications using more than one platform are included.

Figures
Figures

Table 2 .
Use case for chatbots in oncology, with examples of current specific applications or proposed designs.