黑料网

Autism spectrum disorder; Artificial Intelligence; Healthcare; Algorithms; Human intelligence

Availability of massive and expanding quantities of digitized healthcare data, together with advances in computational and storage technologies and machine learning approaches, have opened up new possibilities for AI in healthcare [9]. A multidisciplinary branch of computer science, AI leverages computers to develop systems or algorithms capable of undertaking or partaking in tasks that would otherwise rely completely on human intelligence [10]. Within the healthcare sector strong economic investment in AI-powered technologies [11] has contributed to an exponential growth in topical research [12]. This focus has begun to filter into clinical practice with over 160 AI-powered devices having already been granted regulatory clearance, approval or marketing authorization by the Food and Drug Administration (FDA) [13].

With capabilities to extract clinically meaningful insights from rapidly increasing volumes of healthcare data that have exceeded human analytic capacity [14], AI offers multiple opportunities to augment healthcare practice. Machine learning, a subtype of AI where algorithms are applied to large datasets to look for patterns, can be used to create models that encapsulate those patterns to help predict outcomes [15,16]. Recent reviews have highlighted the promise of machine learning to enhance risk prediction, streamline some diagnostic practices, and support a more data-driven approach to clinical decisionmaking [17,18]. AI-enabled tools may improve accuracy and efficiency of diagnosis, leading to treatment or intervention, and scalability by quickly managing repetitive processes, storing and handling large amounts of data, and providing support for diagnostic or treatment decisions that may reduce the probability for mistakes [19,20]. Use of deep neural networks to augment interpretation of medical scans and other image based data, is one AI application that has received considerable attention [14,21]. A number of recent studies have also explored the therapeutic potential of AI-powered technologies, as well as its capacity to streamline administrative tasks [22]. For example, natural language processing based AI solutions are being leveraged to automate clinical documentation [23,24]. Such approaches show potential to improve workplace efficiency and increase the time clinicians can dedicate to patient care.

A number of promising ASD specific AI-based approaches have been described in the recent literature. The following section of this review highlights key challenges within the ASD diagnostic and therapeutic space and considers the potential of some of these technologies to augment care pathways.

AI in the ASD diagnostic landscape

Current diagnostic challenges

While ASD can reliably be diagnosed as young as 18 months of age, diagnostic challenges and workforce capacity issues in the U.S. are leading to prolonged wait times and delayed initiation of ASD specific treatments. The current average age of ASD diagnosis remains high at over four years [1,25], with roughly 27% of children still undiagnosed at age 8 [8]. While equity of access to diagnostic assessments is improving [1], certain groups such as girls, children who are non-white, of lower socio-economic status, or rural residing, have been noted in the literature to be more often un-diagnosed, mis-diagnosed, or receive a delayed diagnosis [7,26]. ASD evaluation is based on behavioral observations, highlighting the need for more objective methods for ASD assessment with the potential to better understand heterogeneity and identify potential phenotypes that could guide treatment. In response, calls have been made to develop a more data-driven and equitable approach to ASD diagnosis.

ASD has traditionally been diagnosed in specialist settings in the U.S. The dramatic rise in the number of children requiring ASD evaluations has exceeded specialist capacity, however, and resulted in prolonged waits for specialist evaluations. To help facilitate access to early intervention services, the recently updated 2020 American Academy of Pediatrics (AAP) clinical report encourages general pediatricians to make an initial diagnosis of ASD within the medical home for those not requiring specialist referrals [27]. Increasing primary care diagnostic capacity could reduce some of the pressure placed on specialist services and potentially streamline diagnosis and treatment initiation. Unfortunately, current ASD diagnostic tools can be difficult to use in primary care settings as they are time intensive and often require specialist training to administer [28]. They have also only been clinically validated for in-person use, presenting an additional challenge in the context of the ongoing COVID-19 pandemic [29]. Novel approaches that could be administered remotely or within the primary care setting, could potentially decrease some of these diagnostic challenges and workforce capacity issues.

Emerging AI-based innovations

There are a growing number of studies exploring the potential utility of AI in ASD screening and diagnosis. Within the brain magnetic resonance imaging study space, a recent systematic review and meta-analysis identified 43 studies investigating machine learning for ASD diagnosis [30]. Diagnostic accuracy appeared highest for the structural magnetic resonance imaging sub-group of studies, however, multiple methodological limitations were noted across studies. Further robustly designed follow-up trials are needed to clarify the utility of these approaches.

AI is also being used to mine electronic medical records and uncover ASD comorbidity patterns which could enhance screening practices. One novel ASD prediction approach [31] developed digital bio-markers based on the medical histories of patients aged 6 years and under. Data from a commercial claims and encounters database along with data from de-identified diagnostic records from a separate large medical center database was used to train and validate the model. For children over two years of age, the study investigators were able to leverage this model to identify ASD high risk with an area under the receiver operating characteristic above 80%. The study authors note that the autism comorbid risk score (ACoR) they were able to estimate from this work, displays a superior predictive performance to commonly used questionnaire-based screeners and is potentially less biased across demographic groups. An ACoR is an estimate of the likelihood of later ASD diagnosis based on the comorbidity history. An earlier topical study [32] similarly drew on electronic medical records (20K+ patients) and listed medical comorbidities to develop algorithms to detect clinically unique ASD subgroups.

Computer vision AI technology is also being applied in the ASD screening space. Recently, researchers have used deep learning facial image analysis to propose a more objective ASD screening solution [33]. This work draws on noted phenotypic facial differences between typically developing children and children with ASD [34]. The resulting model had 95% classification accuracy. The authors highlight the potential of such an approach to address the subjectivity of current screening practices. Further model training on race-specific datasets could potentially also address some of the racial biases [26] apparent in current practice. A mobile device app designed to capture and distinguish between the eye-gaze patterns of typically developing toddlers and toddlers with ASD has also recently been developed [35]. Designed for use in pediatric primary care settings, the app uses differences detected in the gaze patterns of children with ASD, including poorer coordination of gaze with speech, and reduced gaze response to social stimulus. Application of AI to kinematic features has also shown some potential within the screening and diagnosis space (for a review see [36]). One topical study [37] used a supervised machine learning approach to differentiate between typically developing children and children with severe levels of functional impairment related to ASD based on seven upper limb movement features. Analysis of speech prosody with machine learning techniques has also been explored with some success in ASD screening research [38].

A prescription AI-based Software as a Medical Device [39], prospectively validated, has been developed to aid in the diagnosis of ASD in primary care settings. Following strong clinical trial results, the Device was granted FDA marketing authorization in 2021 [40]. The Device is a diagnosis aid, rather than a standalone diagnostic tool. It leverages a machine learning algorithm developed using patient record data from thousands of children with diverse conditions, presentations, and comorbidities who were either diagnosed with ASD or confirmed not to have ASD based on standardized diagnostic tools and representing both genders across the supported age range [41-45]. Questionnaires combining data and clinical experience were developed to identify the maximally predictive behavioral features of ASD based on the categories of social communication, verbal communication, facial expression, and repetitive behaviors as outlined by DSM-5 criteria. The Device combines 3 independent inputs, all of which can be completed remotely (see figure 1). Uploaded information is evaluated based on predictive features that are most indicative of ASD and one of 3 outputs are provided for the primary care provider to use in conjunction with their clinical judgment: ASD positive, ASD negative, or indeterminate/no result. The latter output acts as a risk control measure [46,47] when information is insufficiently granular to make a diagnostic recommendation with confidence (Figure 1).

Figure 1: Graphical representation of the Device and its major components.

Approaches such as those highlighted above illustrate the breadth of novel AI-based work being undertaken in the ASD screening and diagnostic space. Next, we briefly review key challenges and some potential AI-based solutions within the ASD intervention and therapeutics space.

AI in the ASD treatment landscape

Current treatment challenges

Tailored interventions during the critical neurodevelopmental window can enhance long term outcomes, including gains in cognitive [48,49] and adaptive functioning [4,50], receptive and expressive language use [51], and social skills [3,52]. Despite recognition of the value of early intervention, only one in three children with ASD in the U.S. is thought to be receiving standard of care treatment [53]. Reasons for these gaps include inconsistent access to care providers, especially across regional and remote areas [54], an insufficient number of trained care providers to meet the growing need for ASD intervention, and high out-of-pocket treatment costs, in some cases up to $80,000 per year [55]. Difficulties accessing in-person care providers due to the ongoing Covid-19 pandemic has further curtailed treatment access for some families.

Emerging AI-based innovations

A number of promising AI-based ASD treatment innovations are explored in the literature, including the use of robot assisted therapy [56]. Studies have variously utilized autonomous, partially autonomous, and non-autonomous robot models [57]. Social robottherapy studies have reported improvements in eye contact, emotional recognition and expression, imitation, shared attention to a common object, turn-taking, motor skills and learning behaviors [57-60]. Much of this research is still experimental, however, and large clinical validation studies are needed to quantify the real world potential of such technologies in ASD treatment [61].

Preventing meltdowns using AI technology is another interesting area of ASD intervention research. Behavioral precedents to challenging behavior in children with ASD were identified in one study using deep learning techniques [62]. Researchers then developed a caregiver alert mechanism. This mechanism was designed to provide caregivers with timely warnings so they could potentially intervene prior to behavior escalation. A more recent study [63] similarly sought to develop meltdown prevention signals via AI computer-vision techniques which facilitated real-time facial expression monitoring in children with ASD.

In other research, neural networks have been leveraged to predict how children with ASD will respond to behavioral therapy [64]. This approach was used to explore the relationship between therapy, supervision intensity, age, and gender, on mastery of learning outcomes. Machine learning models have also been built to provide predictive recommendations for the most suitable types of technological treatment interventions for children with ASD [65]. Treatment recommendations were made based on singular or combination symptom patterns. Recent research [66] has also explored the use of an AI-augmented learning and applied behavior analytics platform to personalize ASD intervention. Study authors highlight the potential of such approaches to enhance data-driven clinical decisionmaking, improve intervention efficacy and streamline care delivery.

Clinical adoption challenges

Despite advances in utilizing AI in healthcare, widespread clinical adoption has been limited. Barriers to broader adoption include difficulties integrating such technologies with existing electronic medical record systems, and ethical and legal concerns [67-69]. Regulatory frameworks that account for the differences between AIbased devices and other types of medical devices [70] also require further development. In addition, many AI models require large datasets to train on. However, combining multiple datasets presents technical challenges, and concerns over data privacy and ownership have been raised [71]. The quality of data supplied to a machine learning algorithm is also of critical importance; data bias or imbalance can limit model generalizability and perpetuate pre-existing inequities if not accounted for [22,72]. Questions about a lack of transparency in certain types of AI algorithms and the implications of this within the context of clinical decision-making, have also been raised [14]. Several different approaches to algorithmic explainability and trustworthiness have been proposed in response to this concern [14,73]. Clinician and patient focused AI education is also required prior to broad deployment of these technologies within care pathways. Currently, however, AI education in medical training is patchwork and insufficient [74-77]. Some patients have also expressed concern over data-security and safety standards for AI-based health technologies [78].

Rapidly rising demand for ASD evaluations and treatment [1] has strained workforce capacity and led to suboptimal care for some children. Understaffing, long specialist wait lists, lengthy and complex assessment processes, demographic biases, and limited primary care evaluation capacity have all contributed to delays in diagnosis and treatment initiation [79]. Reduced in-person care options during the ongoing Covid-19 pandemic has further exacerbated access challenges. AI-based innovation in the ASD treatment and diagnostic space shows potential to help address some of these practice challenges. However, additional research is needed to comprehensively clarify the utility and efficacy of many of these approaches. Challenges relating to regulation, data selection and integration, algorithmic transparency and patient and clinician preparedness will all need to be addressed prior to widespread clinical adoption. While much of this research is still in its infancy, the promise of AI in the field of ASD has become clear. As sophistication of machine learning approaches increase, and the volume of digitized medical information and computational power and storage capabilities continue to expand, the impact of intelligent technology on the healthcare landscape will likely accelerate at a rapid pace [10,80].

Conflicts of interest

Dr Shannon, Mr. Abbas, Dr Chettiath, Dr. Salomon and Dr. Taraman are employees of Cognoa and have Cognoa stock options. Dr. Taraman additionally receives consulting fees for Cognito Therapeutics, volunteers as a board member of the AAP - OC chapter and AAP – California, is a paid advisor for MI10 LLC, and owns stock for NTX, Inc., and HandzIn.

1. Maenner MJ, Shaw KA, Bakian AV (2021) . MMWR 70:1-16.

  

2. Estes A, Munson J, Rogers SJ, Greenson J, Winter, J et al. (2015) . J Am Acad Child Adolesc. Psychiatry 54(7):580-587.

      Crossref

3. MacDonald R, Parry-Cruwys D, Dupere S, Ahearn W (2014) . Res Dev Disabil 35:3632-3644.

      Crossref

4. Smith T, Klorman R and Mruzek DW (2015) . J Abnorm Child Psychol 43:1271–1282.

      Crossref

5. Delobel-Ayoub M, Ehlinger V, Klapouszczak D, Maffre T, Raynaud J-P, et al. (2015) . PloS one 10:e0141964.

     

6. Oswald DP, Haworth SM, Mackenzie BK, Willis JH (2017) . Focus on Autism and Other Developmental Disabilities. SAGE Publications 32(2):152–160.

  

7. Wiggins LD, Durkin M, Esler A, Lee L-C, Zahorodny W, et al. . Autism Res 13:464-473.

     

8. Shattuck PT, Durkin M, Maenner M, Newschaffer C, Mandell DS, et al. (2009) J Am Acad Child Adoles Psychiatry 48:474-483.

  

9. Shu L-Q, Sun Y-K, Tan L-H, Shu Q, Chang AC (2019) World J Pediatr 15:105-108.

    

10. Chang AC. . ISBN:978-0-12-823337-5

                

11. Buch VH, Ahmed I, Maruthappu M (2018) Br J Gen Pract 68:143-144.

 

12. Kokol P, Završnik J, Vošner HB (2018) Pediatr Dimens 2:1-5.

     

13. Katznelson G, Gerke S (2021) Adv Health Sci Educ Theory Pract 26:1447-1458.

     

14. Topol EJ (2019) Nat Med 25:44-56.

     

15. Nichols JA, Herbert Chan HW, Baker MAB (2019) Biophysical Reviews 11: p. 111-118.

    

16. Shah NH, Milstein A, Bagley P, Steven C (2019) JAMA 322:1351-1352.

     

17. Chahal D, Byrne MF (2020) . Gastrointest Endosc 92: 813-820.

  

18. Greenhill AT, Edmunds BR (2020) . TIGE 22: 85-89.

  

19. Ahuja AS (2019) . PeerJ 7: e7702.

     

20. Rajkomar A, Oren E, Chen K, Dai AM, Hajaj N, et al (2018) NPJ Digital Medicine 1:18.

    

21. Bressem KK, Adams LC, Gaudin RA, Tröltzsch D, Hamm B, et al (2020) Bioinformatics 36:5255-5261.

     

22. Greenfield D (2019)

 

23. Coiera E, Kocaballi B, Halamka J, Laranjo L (2018) . NPJ digital medicine Nature Publishing Group 1:1-5.

  

24. Kreimeyer K, Foster M, Pandey A, Arya N, Halford G, et al. (2017) . J Biomed Inform 73:14-29.

    

25. Van’t Hof M, Tisseur C, van Berckelear-Onnes I, van Nieuwenhuyzen A, Daniels AM, et al. (2019) . Autism SAGE Publications 25: 862-873.

      

26. Angell AM, Empey A, Zuckerman KE (2018) . Int Rev Res Dev Disabil 55:145-180.

     

27. Hyman SL, Levy SE, Myers SM (2019) 145: e20193447.

     

28. Kaufman NK. (2020) . Appl Neuropsychol Child 27:1-12.

  

29. Dahiya AV, DeLucia E, McDonnell CG, Scarpa A (2021) . Res Dev Disabil 109:103852.

     

30. Moon SJ, Hwang J, Kana R, Torous J, Kim JW (2019) . JMIR ment health 6:e14108.

    

31. Onishchenko Dmytro, Huang Yi, van Horne James, Smith Peter J, Msall Michael E, et al. (2021) . Sci Adv 7:eabf0354.

    

32. Lingren T, Chen P, Bochenek J, Doshi-Velez F, Manning-Courtney P, et al. (2016) . PloS one 11:e0159621.

    

33. Lu A, Perkowski M (2021) Brain Sci 11:1446.

     

34. Obafemi-Ajayi T, Miles JH, Takahashi TN, Qi W, Aldridge K, et al. . J Autism Dev Disord 45:1302–1317.

    

35. Chang Z, Di Martino JM, Aiello R, Baker J, Carpenter K, et al. . JAMA pediatr 175: 827-836.

     

36. Ghosh T, Al Banna MH, Rahman MS, Kaiser MS, Mahmud M, et al. . Sustainable Cities and Society 74:103189.

     

37. Crippa A, Salvatore C, Perego P, Forti S, Nobile M, et al (2015) . J Autism Dev Disord 45:2146-2156.

    

38. Nakai Y, Takiguchi T, Matsui G, Yamaoka N, Takada S (2017) Percept Mot Skills 124:961-973.

     

40. Cognoa. (2020) .   
41. Abbas H, Garberson F, Liu-Mayo S, Glover E (2020) . Sci Rep 10:1–8.

      Crossref

42. Abbas H, Garberson F, Glover E, Wall DP (2018) . J Am Med Inform Assoc 25:1000–1007.

     

43. Duda M, Haber N, Daniels J, Wall DP (2017) . Transl Psychiatry 7:1133–1133.

     

44. Kosmicki JA, Sochat V, Duda M, Wall DP (2015) . Transl Psychiatry 5:514–514.

    

45. Levy S, Duda M, Haber N, Wall DP (2017) . Mol Autism 8:1–17.

  

46. Cortes C, DeSalvo G, Gentile C, Mohri M (2016) . Int Conf Mach Learn 9:1059–1067.

 

47. Kompa B, Snoek J, Beam AL (2021) . NPJ Digit Med 4:1–6.

  

48. Itzchak E, Zachor DA (2011) Res Autism Spectr Disord 5:345–350.

  

49. Flanagan HE, Perry A, Freeman NL (2012) . Res Autism Spectr Disord 6:673–682.

    

50. Dawson G, Rogers S, Munson J, Smith M (2010) l. Pediatrics 125:17-19.

    

51. Vivanti G, Dissanayake C (2016) . J Autism Dev Disord 46:2441-2449. Indexed at       Crossref      
52. Mazurek MO, Kanne SM, Miles JH (2012) . Research in Autism Spectrum Disorders 6:535-545.  

  

53. Penev Y, Dunlap K, Husic A, Hou C, Washington P, et al. (2021) . Appl Clin Inform 12:1030-1040.

      Crossref

54. Ning M, Daniels J, Schwartz J, Dunlap K, Washington P, et al. (2019) . J Med Internet Res 21:e13094.

     Crossref  

55. Cidav Z, Munson J, Estes A, Dawson G, Rogers S, et al. (2017) J Am Acad Child Adolesc Psychiatry 56:777-783.

        Crossref  

56. Palestra G, De Carolis B, Esposito F (2017) Artificial Intelligence for Robot-Assisted Treatment of Autism17-24.
57. Saleh MA, Hanapiah FA, Hashim H (2021) 16:580-602.  

      

58. Admoni H, Scassellati B (2017) . J hum robot interact 6:25-63.
59. Palestra G, Varni G, Chetouani M, Esposito F (2016) n. InProceedings of the International Workshop on Social Learning and Multimodal Interaction for Designing Artificial Agents 1-6.

   

60. Chevalier P, Martin JC, Isableu B, Bazile C, Iacob DO, et al. (2016) 849-854.

       Crossref

61. Jaliaawala MS, Khan RA (2020) . Artificial Intelligence Review 53:1039-1069.
62. Patnam VS, George FT, George K, Verma A (2017) 391-396.

    Crossref

63. Silva JL, Oliveira I, Topolniak Z, Alvarez AB (2021) (SCLA) 1-6.  

    Crossref

64. Linstead E, German R, Dixon D, Granpeesheh D, Novack M, et al. (2015) (ICMLA) 414-418.

    Crossref

65. Akhtar N, Feeney M (2020) (ISSC) 1-6.

    Crossref

66. Ghafghazi S, Carnett A, Neely L, Das A, Rad P (2021) . IEEE Systems, Man, and Cybernetics Magazine 7:4-12.

Crossref

67. Fiske A, Henningsen P, Buyx A (2019) . J Med Internet Res 21:e13216.

      Crossref

68. Davenport T, Kalakota R (2019) . Future Hosp J 6:94.

     

69. Floridi L, Cowls J, Beltrametti M, Chatila R, Chazerand P, et al. (2021) . In Ethics, Governance, and Policies in Artificial Intelligence 19-39.

      Cross Ref

70. Kagiyama N, Shrestha S, Farjo PD, Sengupta PP (2019) . J Am Heart Assoc 8:e012788.

      

71. He J, Baxter SL, Xu J, Xu J, Zhou X, et al. (2019) . Nat Med.Nature Publishing Group 25: 30–36.

       Crossref

72. Zhang H, Lu AX, Abdalla M, McDermott M, Ghassemi M (2020) Proceedings of the ACM Conference on Health, Inference, and Learning: 110–120.

   

73. Ghassemi M, Oakden-Rayner L, Beam AL (2021) . The Lancet Digital Health 3: e745-e750.

   

74. Paranjape K, Schinkel M, Panday RN, Car J, Nanayakkara P. (2019) . JMIR medical education JMIR Publications Inc 5: e16048.

        Crossref

75. Banerjee M, Chiew D, Patel KT, Johns I, Chappell D, et al. (2021) . BMC med educ Bio Med Central 21: 1–10.

   

76. Dos Santos DP, Giese D, Brodehl S, Chon SH, Staab W, et al. (2019) . European Radiology 29: 1640–1646.

  

77. Sit C, Srinivasan R, Amlani A, Muthuswamy K, Azam A, et al. (2020) . Insights into imaging Springer 11: 1–6.

   

78. Richardson JP, Smith C, Curtis S, Watson S, Zhu X, et al. (2021) . npj Digit Med 4: 1-6.

   

79. Gordon-Lipkin E, Foster J, Peacock G. (2016) . Pediatr Clin North Am 63: 851–859.

    Crossref

80. Wartman SA, Combs CD. (2018) Medical education must move from the information age to the age of artificial intelligence. Academic Medicine LWW 93: 1107-1109.