Article Text

Prospective evaluation of artificial intelligence (AI) applications for use in cancer pathways following diagnosis: a systematic review
  1. Sheba Macheka1,
  2. Peng Yun Ng1,
  3. Ophira Ginsburg2,
  4. Andrew Hope3,
  5. Richard Sullivan1 and
  6. Ajay Aggarwal1,4
  1. 1Institute of Cancer Policy, King's College London Faculty of Life Sciences & Medicine, London, UK
  2. 2National Cancer Institute Center for Global Health, Bethesda, Maryland, USA
  3. 3Radiation Oncology, Princess Margaret Hospital Cancer Centre, Toronto, Ontario, Canada
  4. 4Dept of Health Services Research & Policy, London School of Hygiene and Tropical Medicine, London, UK
  1. Correspondence to Professor Richard Sullivan; richard.sullivan{at}kcl.ac.uk

Abstract

The role of artificial intelligence (AI) in cancer care has evolved in the face of ageing population, workforce shortages and technological advancement. Despite recent uptake in AI research and adoption, the extent to which it improves quality, efficiency and equity of care beyond cancer diagnostics is uncertain to date. Henceforth, the objective of our systematic review is to assess the clinical readiness and deployability of AI through evaluation of prospective studies of AI in cancer care following diagnosis.

We undertook a systematic review to determine the types of AI involved and their respective outcomes. A PubMed and Web of Science search between 1 January 2013 and 1 May 2023 identified 15 articles detailing prospective evaluation of AI in postdiagnostic cancer pathway. We appraised all studies using Risk of Bias Assessment of Randomised Controlled Trials and Risk of Bias In Non-randomised Studies-of Interventions quality assessment tools, as well as implementational analysis concerning time, cost and resource, to ascertain the quality of clinical evidence and real-world feasibility of AI.

The results revealed that the majority of AI oncological research remained experimental without prospective clinical validation or deployment. Most studies failed to establish clinical validity and to translate measured AI efficacy into beneficial clinical outcomes. AI research are limited by lack of research standardisation and health system interoperability. Furthermore, implementational analysis and equity considerations of AI were largely missing.

To overcome the triad of low-level clinical evidence, efficacy-outcome gap and incompatible research ecosystem for AI, future work should focus on multicollaborative AI implementation research designed and conducted in accordance with up-to-date research standards and local health systems.

  • radiation oncology
  • chemotherapy
  • palliative care
  • surgical oncology

Data availability statement

All data relevant to the study are included in the article or uploaded as supplementary information. Not applicable.

https://creativecommons.org/licenses/by/4.0/

This is an open access article distributed in accordance with the Creative Commons Attribution 4.0 Unported (CC BY 4.0) license, which permits others to copy, redistribute, remix, transform and build upon this work for any purpose, provided the original work is properly cited, a link to the licence is given, and indication of whether changes were made. See: https://creativecommons.org/licenses/by/4.0/.

Statistics from Altmetric.com

Request Permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.

Introduction

Cancer care is becoming more complex with demographically ageing populations, rising socioeconomic inequalities and the rapid development of novel technologies for treatment.1 2 This complexity coupled with healthcare workforce shortages and infrastructure deficits have created significant opportunities for artificial intelligence (AI) technologies to reshape cancer care across a range of domains and the patient pathway.3

To date, the impact of AI technology is most prominent in cancer diagnostics, particularly radiology.4 AI-based breast screening system has achieved non-inferior performance in interpreting mammograms when compared with expert clinicians in a real-world environment.5 Paige Prostate, an AI software that improves the accuracy and efficiency of prostate biopsy diagnosis, has also received Food and Drug Administration approval in the USA.6 In the field of cancer treatment planning, AI technology has also made significant inroads. OSAIRIS, an open-source AI in medical image analysis, was piloted in a UK hospital and proven to significantly shorten the time required for radiotherapy planning.7 Similar progress has been observed in other areas, such as patient monitoring, precision oncology, behavioural modification and treatment response prediction. For example, machine learning AI models have been incorporated into the evaluation of cell-free DNA advancing the development of liquid biopsy by increasing detection rate and improving monitoring of cancer.8

Nonetheless, progress in applying AI in oncology and healthcare has been tempered by uncertainties regarding the feasibility of AI integration into routine clinical pathways and the extent to which it actually improves the quality, efficiency and equity of cancer care.9 This is, in part, based on concerns regarding the lack of large-scale prospective evaluation of AI algorithms in diverse clinical settings to establish clinical robustness, resource savings and budget impact. AI algorithms are prone to biases that can negatively impact their performance, particularly when they are trained on inadequate, heterogeneous and retrospective datasets.10–12 The case of IBM Watson Health’s cancer AI algorithm is a relevant example, as many of the treatment options initially recommended for patients with cancer have been found to contain significant errors.13 Although it is currently clinically deployed in multiple countires, namely Brazil, China, India, South Korea and Mexico, the most recent concordance study in 2019 reveal varying level of disconcordance between AI-derived recommendations and standard.14

Furthermore, AI faces implementation barriers across very different health ecosystems around data security, antiquated or dynamic regulatory guidelines, administrative burden and lack of research standardisation.15 16 The potential for inequalities in healthcare delivery through the deployment of AI interventions has also been raised by WHO, as the design, development and delivery of AI runs the risk of widening existing disparities or entrenching biases.17

To the best of our knowledge, no systematic attempt has been made to assess the readiness and deployability of AI in oncology, beyond cancer diagnostics. Our systematic review focused specifically on evaluating the level of evidence for new AI solutions, specifically identifying studies that have sought prospective evaluation of AI tools in the postdiagnosis cancer pathway. The aim was to gain insight into the research landscape of AI, specifically the breadth and quality of studies evaluating AI algorithms to support the cancer pathway. In doing so, we can identify potential barriers to implementation and future research needs.

Methods

The systematic review was designed using the Preferred Reporting Items for Systematic reviews and Meta-analyses. Studies published between 1 January 2013 and 1 May 2023 were searched on two online databases using PubMed (inclusive of Medline) and Embase. The following search terms were used to identify suitable publications: “Artificial Intelligence or Machine Learning or Deep Learning or Neural Network and Cancer or Tumour or Malignant”. The full search strategy used is available in the online supplemental data (online supplemental appendix A).

Supplemental material

Inclusion criteria

Studies in the postdiagnostic care pathway, evaluating validated AI algorithms to assess the efficacy or quality of the AI and/or its efficiency in patient workflows; adult solid organ malignancies or multiple tumour sites (this may include haematological malignancies), written in English and focused on human adults. Publications were accepted if they were prospective in nature, including (phase I–IV clinical trials), case-control studies and observational studies.

Exclusion criteria

All publications that are retrospective or focused solely on haematological or paediatric malignancies are excluded. Review articles, letters, abstracts, conferences proceedings, editorials, preclinical studies, trial protocols and all studies published in non-English language are also excluded. Any studies in the development or validation phase and those training or evaluating an AI tool on retrospective datasets from the same institution were excluded.

Data selection

The titles and abstracts were assessed. All potential abstracts were identified for full-text review. The studies were initially selected by SM with AA to check and assess the excluded studies. SM and PYN extracted data from each study, assessing its quality and any uncertainty was reviewed by AA and RS.

Data extraction

The data included in the extraction are as follows:

  • Location of studies.

  • Characteristics of study (funding, setting, research design and sample size).

  • Tumour site.

  • Purpose of AI and component of cancer care pathway where AI was applied.

  • Outcome of evaluation.

  • Time, cost and resource use analysis.

Data analysis

PYN conducted a risk of bias quality assessment for all randomised controlled trials and non-randomised interventional studies, using the Risk of Bias Assessment of Randomised Controlled Trials (ROB-2) and Risk of Bias In Non-randomised Studies-of Interventions (ROBIN-I) quality assessment tools, respectively.

Results

Search strategy

12457 publications were initially identified in the PubMed (inclusive of Medline) and Embase database. 32 studies were selected after the abstracts and titles were screened. Of these, 17 were excluded because of (1) being in the development and validation phase, (2) not directly investigating an AI tool, (3) preclinical phase studies or (4) clinical trial protocols. 15 publications fulfilled the search criteria and are summarised in online supplemental table S1 (online supplemental appendix B).

The search strategy is illustrated in figure 1.

Figure 1

The Preferred Reporting Items for Systematic reviews and Meta-analyses flow chart of identification for articles for inclusion.

Tumour type, location, type of studies

The majority of studies focused on two or more tumour sites (n=8),18–25 with a particular emphasis on breast (n=8)18–20 23 24 26 27 and gastrointestinal cancer (n=8).18–24 28 Other cancer types included prostate (n=5),20 22 25 29 30 gynaecological (n=4)20 21 23 24 as well as thyroid and head and neck (n=5).18 19 24 25 31

Most studies were published in the USA (n=9).18–24 26 32 Other published countries included Canada (n=3),25 29 30 China (n=2)27 28 and South Korea (n=1).31

All the included studies were prospective and described AI in the postdiagnostic cancer care pathway. Almost half were randomised control trials (n=7),18–21 26 29 31 while the other half were observational studies (n=8).22–25 27 28 30 32

Types of AI investigated and clinical pathways

Clinician behaviour

Six studies focused on the use of AI to influence the patients’ or clinicians’ behaviour.19 20 23 26 27 31 Two studies identified whether using a machine learning algorithm predicting 180-day mortality and identifying high-risk patients with different cancer types, influences a clinicians’ decision to instigate serious illness conversations (SIC) and advanced care planning. In both studies, there was a significant increase in SIC from 3.4% to 13.5%.19 23 In addition, there was a decrease in the rates of systemic treatments used in patients approaching the end of life from 10.4% to 7.5%.19 23 Both studies were limited due to single healthcare electronic record. A further study assessed whether the artificial Clinical Decision Support System (CDSS) can change clinical treatment decisions in patients with breast cancer.27 The CDSS is an AI-based software that can make treatment recommendations to oncologists, based on clinical data drawn from clinical records.27 This study showed that treatment decisions changes in 5% of patients and more likely in hormone-positive or stage 4 disease (first-line treatment).27

Patient behaviour

AI can impact a patients’ behaviour, particularly with lifestyle measures such as exercise.20 26 A pilot study conducted over a 4-week period illustrated that an AI-based voice-coaching programme increased the average step-count by 3568.9 steps/day for overweight, physically inactive cancer survivors, compared with the control group 2160.6 steps/day (p<0.05).20 This study was limited by the small number of participants (n=42) and short follow-up period of 4 weeks.20 A second study assessed whether machine learning can provide accurate estimates of physical activity (PA), as there tends to be discrepancies between self-reporting and accelerometer data.26 This study showed that self-report and machine learning provided similar PA estimates at baseline (mean difference=11.5 min/day) and the mean difference of PA change for the cut-point versus machine learning methods was 5.1 min/day for intervention group and 2.9 in controls.26 Lastly, AI can be used to improve a patients’ understanding of their disease and improve the informed consent process for thyroid surgery.31 A deep neural network was used to design a personalised three-dimensional thyroid model.31 The group in the experimental arm showed a better understanding of their disease process, as well as the benefits and risks of thyroid surgery.31 However, the study was not blinded, which may have introduced an element of bias.31

Survival

Two studies assessed whether a machine learning can accurately predict prognosis in patients with advanced cancer.22 23 The Number of active tumors (“N”), Eastern Cooperative Oncology Group performance status (“E”), albumin (“A”) and primary tumor site (“T”) (NEAT) model provided better accurate prognostic predictions which was statistically significant, compared with experienced oncology physicians and nurses.22 However, its generalisability is limited due to it being single-site study.22 On the contrary, the other study, Manz et al demonstrated the feasibility and clinical validation of a machine learning for real-time short-term prognosis of patients with cancer across 18 centres under a single academic health system.23

Treatment

Four studies assessed AI in radiotherapy and brachytherapy.25 29 30 32 A machine learning brachytherapy treatment planning system for prostate cancer was tested in a prospective clinical trial and showed non-inferiority, in comparison with manual planning for the dose measured at implantation and 30 days.29 In addition, there was a significant reduction in treatment planning time.29 Similarly, a study focused on an AI auto-segmentation radiotherapy planning system for patients with non-small cell lung cancer with nodal involvement, showed a 65% reduction in segmentation time (p<0.0001).32 However, the major limitations were the lack of PET scan imaging for the radiotherapy planning, which is a common clinical application available for radiotherapy planning in lung cancer cases.32 A study conducted in China developed and tested an AI RAdioPathomics Integrated preDiction System.28 This tool predicted complete response in patient with locally advanced rectal cancer undergoing neoadjuvant chemoradiotherapy, based on pretreatment radiopathomic images with high accuracy.28 However, no demographic details were input into the model, which may improve its performance. A Canadian study trialled a fully workflow-integrated, machine learning-based radiotherapy planning software for patients with prostate cancer. Its outcome highlighted that AI acceptability by clinicians in real life differed from its retrospective evaluation30 (21% decrease in selection of peer-reviewed quantitatively superior machine learning radiotherapy plans by the clinicians at the simulation vs deployment phase, 92% vs 71%). Lastly, another Canadian study evaluated the performance of implemented deep learning-based auto-segmentation for central nervous system, head and neck and prostate cancer radiotherapy planning into the workflow. It concluded that the deep learning-based auto-segmented plans required minimal subjective (mean editing score ≤2) and objective edits (mean Dice similarity coefficient (DSC) and 95% Hausdorff distance (HD) was ≥0.90 and ≤2.0 mm) and resulted in a positive user experience.25

Within surgery, a trial assessed the use of a machine learning model to predict surgical case duration in gynae-oncology and colorectal surgical cases.21 This machine learning algorithm was better in predicting surgical case times, compared with the surgeons (p<0.03).21 This may enable better allocation of clinical resources and reduce patient waiting times.21 However, its accuracy was dependent on the correct data input and the system could only provide short-term predictions, 24 hours prior to the planned surgery.21

Lastly, a machine learning algorithm was trialled in the USA to identify high-risk patients with different solid organ malignancies undergoing radiotherapy or chemoradiotherapy.24 This algorithm identified patients likely to require acute care during their treatments and suggested twice-weekly on-treatment reviews.24 Twice-weekly evaluation reduced rates of acute care from 22.3% to 12.3% (p=0.02).24 Its main limitation is that it was conducted in a single centre.

Quality assessment of studies

Randomised controlled trials

All studies involving randomised controlled trials were assessed using the ROB-2 quality assessment tool.33 The outcomes of the quality assessment are summarised in table 1.

Table 1

Risk of Bias Assessment of Randomised Controlled Trials (ROB-2)

Four out of seven randomised controlled trials (57.1%) were rated as having a low overall risk of bias across all five domains of risk of bias using the ROB-2 tool.18 20 29 31 Some concerns of risk of bias from effects of assignment to intervention were noted for Strömblad et al,21 due to 9.5% exclusion (n=72/755) postrandomisation.19 21 Manz et al was judged to have an overall high risk of bias specifically for their secondary outcomes end-of-life care and hospice enrolment, as they were missing 16.2% (n=229/1417) and 39.6% (n=569/1417) of outcome data, respectively due to study team’s dependence on hospital cancer registry for data collection and patients’ eligibility to hospices.19 26 Nelson et al was rated as having a high risk of bias due to their per-protocol analysis with 13% missing data (n=44/333), which have an unpredictable effect on the outcome.26

Observational studies

All studies involving non-randomised interventions were assessed using the ROBINS-I quality assessment tool.34 The outcomes of the quality assessment are summarised in table 2.

Table 2

The Risk of Bias In Non-randomised Studies-of Interventions (ROBIN-I) assessment

Half of the non-randomised studies of intervention were rated as having low overall risk of bias, meaning their risk is comparable to that of a well-conducted randomised trial.22 28 30 32 The overall risk of bias of Manz et al was deemed serious due to missing data and subsequent exclusion of patients with missing data from analysis.23 24 Hong et al was judged to have serious risk of bias, as the baseline characteristics of their cohorts were not adjusted to confounders, such as comorbidities, cancer diagnosis, age, gender and other relevant prognostic factors, which could significantly impact their outcomes, specifically emergency care attendance and hospital admission during cancer treatment.24 27 Xu et al looking at the effect of machine learning powered Clinical Decision Support System, suffered from serious risk of bias, as it relied on a small sample size of oncologists, who were not assessed for user acceptability of AI and adjusted for clinical experience prior to intervention.25 27 Wong et al did not blind the reviewers of the contours as to the source of the segmentation (deep learning-based or manual segmentation). The results of the study therefore suffered from a serious risk of bias in favour of the intervention arm.25

Discussion

There is a concern that the clinical impact of AI may be limited or have piecemeal adoption because of the lack of robust evidence demonstrating its efficacy and cost-effectiveness. In this systematic review of all published studies evaluating the role of AI in the postcancer diagnostic pathway, we found only 15 studies that met our study criteria. The studies were predominantly single-centre studies with small sample sizes of ≤50 patients and no studies were conducted in the low-income and middle-income country setting. Overall, most oncology AI research remains at an experimental stage without prospective clinical validation or deployment due to implementation barriers illustrated in figure 2.35 36

Figure 2

Implementation barriers of artificial intelligence (AI).

Low level of evidence

Most studies were in the pilot stage, that is, first time clinical deployment within tight trial parameters with relatively short follow-ups. Consequently, the clinical impact of AI in a real-world setting on integration into the healthcare system remains unknown. In addition, the design and conduct of the studies differed in quality, compromising the robustness of the evidence produced. Half of the studies were randomised controlled trials and the remaining half were observational. Quality assessment for risk of bias, using the ROB-2 and ROBIN-I tools for randomised controlled trials and observational studies, respectively, revealed that there were either at least some concerns or serious risk of bias in nearly half of the studies (n=6, 42.9%), due to unadjusted confounding variables, missing data, selective reporting and inappropriate method of analysis.

Most studies were conducted at a single clinical site (n=9, 64.3%), thus limiting its generalisability and external validity.20–23 25 26 28 30 31 This potentially signalled technical difficulties in designing and developing studies across multiple sites due to a lack of interoperability among hospitals. Both adoption of unified data formats such as Fast Healthcare Interoperability Resources, and consistent clinical coding in electronic healthcare record, as promoted by the Minimal Common Oncology Data Elements in the USA, are necessary to overcome this hurdle to facilitate AI implementation, as demonstrated by multicentre, single health system studies conducted by Manz et al.37 38 The sample size of studies varied but given the context, some were too small (sample size ≤50) to provide sufficient power and precision for clinical deployment of the AI.20 29 30 32 Additional efforts to develop AI-ready data infrastructures rely on ontology approaches, such as the Operational Ontology for Oncology, to standardise real-world data for use in training or testing of novel algorithms.39

Two examples of large scale, high-quality prospective evaluation of AI tools in real-world setting were published by Dembrower et al40 and Lång et al41 to validate AI-supported mammogram screening for breast cancer in Sweden.40 41

Gap between AI efficacy and clinical outcome

First, the measured metrics of the AI tool used in research might not directly translate into clinical benefits. The term ‘AI chasm’ was coined to reflect this phenomenon.42 To illustrate, serious illness conversations (SIC) were used as surrogates for goal-concordant care and less aggressive end-of-life care, assuming that SIC would improve concordance with advanced care plans and influence clinicians’ approach towards end-of-life care, such as avoiding prescription of systemic treatment close to death.18 However, when the same team used different end points with the same AI tool it found no effect on patients’ hospice enrolment, hospital length of stay, inpatient death or end-of-life intensive care unit admission. Thus, a robust clinical evaluation using indicators that are intuitive to clinicians and reflect the quality of care is essential. In the development and evaluation stage, the validity of the indicators should be based on its meaning and relevance to both clinicians and patients, that is, does the end point represent an important measure of quality or outcome relevant to patients.43 A difference in the indicator should also reflect a difference in the quality of care, with a specific direction reflecting better quality.43 Delphi techniques could be adopted to gain consensus among key stakeholders on the most important outcomes.44 An excellent example of AI studies with valid and technically specific indicators is the ongoing prospective ARtificial intelligenCe-based radiotHERapY (ARCHERY) study. This international prospective evaluation has been designed to independently evaluate according to a prespecified protocol the clinical acceptability (based on contouring and dosimetric parameters) of AI-based radiotherapy treatment planning for cervical, head and neck and prostate cancers. Alongside this, time and human resource savings have been estimated as well as a budget impact analysis.45

Second, most AI tools were calibrated using retrospective datasets. Low quality, incomplete and/or discrepant data in the clinical setting can negatively impact its real-world clinical performance.10–12 This is highlighted in the study by Strömblad et al,21 where their algorithm that predicts the duration of surgery was compromised in cases where the actual surgery undertaken in theatre deviated from the presurgical plan.21

Third, all existing studies, except for Hosny et al32 and Wong et al,25 did not explore the experience of AI users, be it healthcare providers or patients.25 32 This lack of insight may not affect the clinical outcome when the AI was only trialled by a small number of users over a short period of time, as demonstrated by Hassoon et al and Xu et al, but it poses uncertainties regarding the user acceptability, feasibility and sustainability of the AI tool in the long run and when it is scaled.20 27 The efficacy gap between retrospective simulation and real-life deployment of AI due to user acceptability was highlighted by McIntosh et al.30 They observed a 21% decrease in clinicians’ selection of AI-generated radiotherapy plans between simulation and treatment deployment even though the plans were deemed superior by blind expert reviews.30 To overcome this shortfall, implementation frameworks such as the Consolidated Framework for Implementation Research should be embedded in early stage research design to assess a range of contextual factors that hinder or facilitate the adoption of AI as a complex intervention, and to inform implementation strategies that may best address contextual determinants, such as clinicians’ bias.46

Fourth, the economics of AI implementation, in terms of time, cost and resources, were largely overlooked. Only two studies by Nicolae et al29 and Hosny et al32 went beyond in silico validation of AI to include a time-saving analysis, while one study by Hong et al24 analysed the potential resource saved using AI.24 29 32 Robust implementation validation should be encouraged in future studies as it is crucial to identify the downstream consequences of AI implementation in health systems, including the structure of care, process of care and workforce planning. This is particularly important in resource-limited countries, as their healthcare systems may not be mature enough to adopt AI tools. Equity consideration of AI interventions was also absent in all studies. To close the translational gap for real-life AI application in different resource settings, frameworks, such as Reach, Effectiveness, Adoption, Implementation and Maintenance, are recommended by WHO to help structure the implementation research.17

Research ecosystem for AI

Our review showed that different methodologies were used to demonstrate efficacy in AI interventions of the same type, that serve the same purpose, indicating that there was a lack of standardisation in the evaluation and validation of AI. Although the area under the curve of the receiver operating characteristics is a popular statistical measure of the performance of machine learning models, it is often not sufficient to prove the clinical efficacy of AI.

Beyond statistical findings, the development environment (ie, clinical setting from which data used to train the model are generated), operational environment (ie, environment in which AI is deployed in including integration with health record system and infrastructure required) and human-AI interaction should be included in the research protocol to allow transparent and more holistic evaluation of an AI intervention, as recommended by the Standard Protocol Items: Recommendations for Interventional Trials - Artificial Intelligence (SPIRIT-AI) extension to the Consolidated Standards of Reporting Trials guidelines.47 Output data and reporting should also be tailored based on the type and purpose of AI. For example, an AI-driven image analytical tool should be reported using a class-activation map to visualise pixels that had the greatest influence on the predicted class.46 For predictive AI models, researchers should refer to the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis-Machine Learning and Standards for Reporting Diagnostic Accuracy Studies-Artificial Intelligence, which were newly developed.48 49 Of the 15 studies included, only 2 studies used their relevant up-to-date guidelines, Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis.23 28

In terms of ethical and regulatory approval, data security is a concern as health data are often sensitive, private and stored in large volume. As an illustration, the AI voice coaching trialled in the study by Hassoon et al20 monitored the users’ day-to-day PA to provide feedback on their advice.20 Confidential data, as such, require safe methods of collection, storage and usage through a secure data server, and battles against data breach and sabotage will require consistent vigilance, investment and legislature protection.

Recommendations

To overcome the triad of insufficient clinical evidence, efficacy-outcome gap and new research ecosystem of AI that currently hinders the implementation of AI in oncology, we propose the recommendations to key stakeholders as mentioned in table 3.

Table 3

Action plans tailored to stakeholders for addressing specific issues

More funding is required overall, but more importantly, a larger proportion of funding should be dedicated to implementation science research of AI. In the UK, the NHS AI Lab, led by the National Institute of Health Research and NHS England, creates a community space for clinicians, data scientists, healthcare providers and regulators. It also facilitates funding and implementation of AI in clinical environment through AI Health and Care Award for AI at different phases of trials.50 Such multidisclipinary collaboration backed by funding should be expanded to bridge the chasm between AI and clinical implementation in routine clinical cancer care.

Strengths and limitations of review

The strengths of our review lie in our comprehensive yet targeted inclusion criteria (specifically prospective studies in postdiagnostic cancer care) and robust methodology triangulated by systematic review and quality assessment. We also provided stakeholder-specific action plans by learning from successful examples available in a wider context.

By design, our review excluded prospective AI studies in cancer diagnostics, which make up the majority of AI research and are the most advanced on the frontier, such as studies by Dembrower et al40 and Lång et al.41 As a result, we were unable to analyse these studies, some of which were deployed in real-world environments and across different healthcare ecosystems. Additionally, previous computational or predictive models that were not classified as AI/machine learning/deep learning/neural networks were excluded, but may have parallels or considerations that should be considered when addressing our recommendations.

Conclusion

AI is a fast-growing technology with immense potential to reshape cancer care and pathways beyond cancer diagnostics. Despite the exponential growth in AI research into postdiagnostic cancer care, only a small fraction of AI tools have undergone prospective clinical evaluation and concerns were highlighted regarding the size of the study, the breadth of participants and the study conduct. Resource, cost-effectiveness and time-saving analyses of AI were largely missing; as were qualitative surveys on user acceptability, feasibility and sustainability of AI.

Future work should focus on multicollaborative AI implementation research co-developed by academics, healthcare providers, commissioners and patients, in accordance with up-to-date research guidelines and local health systems. They should focus on holistically evaluating AI technologies and determining their readiness for safe, feasible and efficient clinical deployment.

Data availability statement

All data relevant to the study are included in the article or uploaded as supplementary information. Not applicable.

Ethics statements

Patient consent for publication

Ethics approval

Not applicable.

References

Supplementary materials

  • Supplementary Data

    This web only file has been produced by the BMJ Publishing Group from an electronic file supplied by the author(s) and has not been edited for content.

Footnotes

  • SM and PYN are joint first authors.

  • X @cut_tt_chase

  • Contributors SM and PYN are joint first authors. RS and AA conceived and designed the study. SM and PYN searched, screened and assessed the publications. SM and PYN extracted, analysed and visualised the data. PYN and SM drafted the manuscript and interpreted the findings. OG, AH, RS and AA reviewed and edited drafts of the manuscript. AA accessed and verified the data. All authors read the manuscript, provided feedback and approved the final version.

  • Funding RS and PYN are funded by Medical Research Council Global Alliance of Chronic Disease Grant ACCI No GACD-025. RS and PYN are also funded by BASO/Rosetrees Research Grant in Cancer Surgery.

  • Competing interests None declared.

  • Patient and public involvement Patients and/or the public were not involved in the design, or conduct, or reporting, or dissemination plans of this research.

  • Provenance and peer review Not commissioned; externally peer reviewed.

  • Supplemental material This content has been supplied by the author(s). It has not been vetted by BMJ Publishing Group Limited (BMJ) and may not have been peer-reviewed. Any opinions or recommendations discussed are solely those of the author(s) and are not endorsed by BMJ. BMJ disclaims all liability and responsibility arising from any reliance placed on the content. Where the content includes any translated material, BMJ does not warrant the accuracy and reliability of the translations (including but not limited to local regulations, clinical guidelines, terminology, drug names and drug dosages), and is not responsible for any error and/or omissions arising from translation and adaptation or otherwise.

Linked Articles