Turning the crank for machine learning: ease, at what expense?

Excitement around the transformative potential of machine learning in health care belies a reliance on deep technical expertise that leaves this technology in the hands of the few. Typically, a practitioner of machine learning undertakes numerous tasks in the process of training and testing a model for classification. The process requires substantial technical knowledge and—perhaps somewhat incongruously—is often both highly detailed and loosely defined. In The Lancet Digital Health, Livia Faes, Siegfried Wagner, and colleagues1Faes L Wagner SK Fuet DJ et al.Automated deep learning design for medical image classification by health-care professionals with no coding experience: a feasibility study.Lancet Digital Health. 2019; 1: 232-242Summary Full Text Full Text PDF Scopus (45) Google Scholar report on their experience of using a service that creates an abstraction from the training and testing process, enabling a professional with no coding experience to build a model that might once have been out of reach. In the study, the authors train models for classifying disease in medical images using Cloud AutoML, a service that requires minimal technical knowledge. Discriminative performance is compared with values previously reported in the academic literature for matching tasks. The study does not claim to introduce new methods for machine learning, but it does highlight the potential of an easily accessible service that could be used by health-care providers. The work is compelling because it suggests that expert-level results in image classification are now achievable by anyone with cursory training. We cautiously share the authors’ optimism that removing obstacles to algorithmic modelling will lead to improvements in patient care, but the risks of bypassing mathematical, statistical, and programming expertise must be emphasised. The use of machine learning methods without in-depth knowledge can result in misleading or outright erroneous results that would cause harm if used to guide the delivery of care. A reliance on simple performance metrics alone does not allow the practitioner to interpret other aspects of model development. For example, a model could demonstrate racial bias by yielding differing results for different subpopulations.2Chen I Johansson FD Sontag D Why is my classifier discriminatory?.ArXiv. 2018; (preprint).https://arxiv.org/pdf/1805.12002.pdfGoogle Scholar Often the data itself introduces its own modelling challenges.3Ghassemi M Naumann T Schulam P Beam AL Chen IY Ranganath R Practical guidance on artificial intelligence for health-care data.Lancet Digital Health. 2019; 1: e157-e159Summary Full Text Full Text PDF Scopus (20) Google Scholar, 4Panch T Mattie H Celi LA The “inconvenient truth” about AI in healthcare.NPJ Digit Med. 2019; 2: 77Crossref PubMed Scopus (18) Google Scholar Artifacts could cause a model to learn spurious rules, for example, such as the skin cancer algorithm that associates suspicious lesions with the surgical skin markers that surround them.5Winkler JK Fink C Toberer F et al.Association between surgical skin markings in dermoscopic images and diagnostic performance of a deep learning convolutional neural network for melanoma recognition.JAMA Dermatol. 2019; (published online Aug 14.)DOI:10.1001/jamadermatol.2019.1735Crossref PubMed Scopus (59) Google Scholar Data quality might drift over time (for example, with changing equipment or operators), confounding an analysis that fails to account for these changes. The models demonstrated in this study perform well on benchmark tasks, but they fail to generalise to external data. In reviewing the performance issues, the authors acknowledge their limited ability to audit the models or the data. Since the costs of misclassification can be high, this lack of transparency is concerning. Considerations of bias and technical rigour should be dominant considerations in health care, an issue that we hope will be addressed as the service develops. The “sharp contrast of the model's discriminative performance” when moving from the “internal” to the “external” testing dermatology dataset leads the authors to conclude that a “small data” approach might be the ultimate use case for automated deep learning software.1Faes L Wagner SK Fuet DJ et al.Automated deep learning design for medical image classification by health-care professionals with no coding experience: a feasibility study.Lancet Digital Health. 2019; 1: 232-242Summary Full Text Full Text PDF Scopus (45) Google Scholar The approach would involve researchers and clinicians training models within their own institutions for “a specific geographical patient population that a given clinic might encounter”.1Faes L Wagner SK Fuet DJ et al.Automated deep learning design for medical image classification by health-care professionals with no coding experience: a feasibility study.Lancet Digital Health. 2019; 1: 232-242Summary Full Text Full Text PDF Scopus (45) Google Scholar While the simplicity of this approach is appealing, caution is clearly needed. What is the fate of the patient from outside this geographic population, previously unseen by the model? We welcome efforts to improve the ease of developing models in health care, but it is important that governance, ethics, and technical oversight are allowed to keep up. Faes, Wagner and colleagues conclude that regulatory guidelines are needed for both medical deep learning and clinical implementation of these models before they might be used in clinical practice. We absolutely agree. A wider discussion about the ethics of training and deploying machine learning models in routine clinical practice—involving multiple disciplines spanning the clinical and computing worlds—must ensue. For now, machine learning in health care should remain collaborative, with experts from across disciplines working together.6Wiens J Saria S Sendak M et al.Do no harm: a roadmap for responsible machine learning for health care.Nature Med. 2019; (published online Aug 19)https://doi.org/10.1038/s41591-019-0548-6Crossref Scopus (101) Google Scholar Meaningful machine learning for health is not just about turning a crank, but it requires the careful and thoughtful application of analytical techniques. This online publication has been corrected. The corrected version first appeared at thelancet.com/digital-health on September 16, 2019 This online publication has been corrected. The corrected version first appeared at thelancet.com/digital-health on September 16, 2019 TJP reports non-financial support from Google Cloud, outside the submitted work. SH reports grants from Philips Research, grants from Doris Dukes Charitable Foundation, outside the submitted work. MG worked previously for Verily, a subsidiary of Alphabet Inc and is supported in part by Microsoft Research. The other authors declare no competing interests. Automated deep learning design for medical image classification by health-care professionals with no coding experience: a feasibility studyAll models, except the automated deep learning model trained on the multilabel classification task of the NIH CXR14 dataset, showed comparable discriminative performance and diagnostic properties to state-of-the-art performing deep learning algorithms. The performance in the external validation study was low. The quality of the open-access datasets (including insufficient information about patient flow and demographics) and the absence of measurement for precision, such as confidence intervals, constituted the major limitations of this study. Full-Text PDF Open AccessCorrection to Lancet Digital Health 2019; 1: e198–99Pollard TJ, Chen I, Wiens J, et al. Turning the crank for machine learning: ease, at what expense? Lancet Digital Health 2019; 1: e198–99—In this Comment, Emily Lindemer's name was incorrectly spelled. This correction has been made as of September 16, 2019. Full-Text PDF Open Access

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.