AI News, A comparison of deep learning performance against health artificial intelligence

Artificial Intelligence is Coming Fast. Are Surgeons Ready?

Once the domain of science fiction and futurism, the technology now translates web pages online, trades stocks on Wall Street and drives automobiles on city streets.

A widely covered paper1 published on Oct. 1 in The Lancet Digital Health concluded that diagnostic systems based on a form of AI called deep learning can match or outperform health care providers in accuracy.

The authors performed a pooled analysis of 69 studies published between Jan. 1, 2012 and June 6, 2019 that developed deep learning models for disease diagnosis based on medical imaging or histopathology. A

The system was trained on 512 patients, and used 87 clinical variables to predict whether the patient would later qualify for ICU care based on 15 criteria such as re-intubation, prolonged hypotension or new-onset arrhythmia.

In comparison, a surgeon, anesthesiologist and intensivist correctly assigned these patients 70 percent, 58 percent and 64 percent of the time respectively.

In one case, Dr. Firriolo said, researchers using AI to predict which whether ICU patients would die within two days discovered that the algorithm was seizing on a single word in the electronic medical records: chaplain.

He has shown that AI can clearly match pathologists in identifying fat globules, and surpasses them in estimating the total fat content of a donor organ.

And because his project is ultimately trying to determine whether donor liver fat content affects the success of transplants, that dataset also needs to include patient outcomes as well. “My

He imagines a future in which AI algorithms will have real-time access to electronic medical record data as it accumulates, and tell physicians when a patient should go to the ICU or receive a transfusion. Faced

For example, the rapid advancement of automated vehicles raises the issue of how to convince the public to adopt AI technology once it’s demonstrably safer than a human operator.

But the reality is that if computers do eventually achieve general intelligence they will inevitably begin to outpace human surgeons, and we should begin to think now about how best to incorporate the technology into surgical practice. “If

Key challenges for delivering clinical impact with artificial intelligence

While existing studies have encompassed very large numbers of patients with extensive benchmarking against expert performance, the vast majority of studies have been retrospective, meaning that they use historically labelled data to train and test algorithms.

The limited number of prospective studies to date include diabetic retinopathy grading [48,49,50], detection of breast cancer metastases in sentinel lymph node biopsies [51, 52], wrist fracture detection [53], colonic polyp detection [28, 54], and detection of congenital cataracts [55].

these include an algorithm to detect childhood cataracts with promising performance in a small prospective study [55] but less accurate performance compared to senior clinicians in a diagnostic RCT [57];

an open, non-blinded randomised trial of an automatic polyp detection algorithm for diagnostic colonoscopy demonstrating a significant increase in detection of diminutive adenomas and hyperplastic polyps [59];

Future studies should aim to use clinical outcomes as trial endpoints to demonstrate longer-term benefit, while recognising that algorithms are likely to result in changes of the sociocultural context or care pathways;

Machine learning studies should aim to follow best practice recommendations, such as the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD), designed to assist the reporting of studies that develop, validate or update a prediction model for either diagnostic or prognostic purposes [63].

In addition, a new version of the TRIPOD statement that is specific to machine learning prediction algorithms (TRIPOD-ML) is in development and will focus on the introduction of machine learning prediction algorithms, establishing methodological and reporting standards for machine learning studies in healthcare [64].

As well as reporting sensitivity and specificity at a selected model operating point (required to turn the continuous model output into discrete decision categories), papers should include information about positive and negative predictive values.

Important factors for consideration include dataset shift, accidentally fitting confounders rather than true signal, propagating unintentional biases in clinical practice, providing algorithms with interpretability, developing reliable measures of model confidence, and the challenge of generalisation to different populations.

Particularly important for EHR algorithms, it is easy to ignore the fact that all input data are generated within a non-stationary environment with shifting patient populations, where clinical and operational practices evolve over time [70].

For instance, in one classic example, a machine learning model did not learn the intrinsic difference between dogs and wolves, but instead learned that wolves are usually pictured standing on snow, while dogs usually appear on grass [72].

Generalisation can be hard due to technical differences between sites (including differences in equipment, coding definitions, EHR systems, and laboratory equipment and assays) as well as variations in local clinical and administrative practices.

Generalisation of model operating points may also prove challenging across new populations, as illustrated in a recent study to detect abnormal chest radiographs, where specificity at a fixed operating point varied widely, from 0.566 to 1.000, across five independent datasets [5].

Proper assessment of real-world clinical performance and generalisation requires appropriately designed external validation involving testing of an AI system using adequately sized datasets collected from institutions other than those that provided the data for model training.

A recent systematic review of studies that evaluated AI algorithms for the diagnostic analysis of medical imaging found that only 6% of 516 eligible published studies performed external validation [77].

Blind spots in machine learning can reflect the worst societal biases, with a risk of unintended or unknown accuracies in minority subgroups, and there is fear over the potential for amplifying biases present in the historical data [78].

In medicine, examples include hospital mortality prediction algorithms with varying accuracy by ethnicity [80] and algorithms that can classify images of benign and malignant moles with accuracy similar to that of board-certified dermatologists [19, 81], but with underperformance on images of lesions in skin of colour due to training on open datasets of predominantly fair skinned patients.

models selected to best represent the majority and not necessarily underrepresented groups), (2) model variance (due to inadequate data from minorities), and (3) outcome noise (the effect of a set of unobserved variables that potentially interacts with model predictions, avoidable by identifying subpopulations to measure additional variables) [80].

A greater awareness of these issues and empowering clinicians to participate critically in system design and development will help guide researchers to ensure that the correct steps are taken to quantify bias before deploying models.

Adoption of unified data formats, such as Fast Healthcare Interoperability Resources [84], offer the potential for better aggregation of data, although improved interoperability does not necessarily fix the problem of inconsistent semantic coding in EHR data [85].

In order to ensure that this technology can reach and benefit patients, it will be important to maintain a focus on clinical applicability and patient outcomes, advance methods for algorithmic interpretability, and achieve a better understanding of human–computer interactions.

Given the combination of the devastating consequences of unacceptable results, the high risk of unquantified bias that is difficult to identify a priori, and the recognised potential for models to use inappropriate confounding variables, explainability enables system verification.

Machine learning methods that build upon a long history of research in traditional symbolic AI techniques to allow for encoding of semantics of data and the use of ontologies to guide the learning process may permit human experts to understand and retrace decision processes more effectively [91, 92].

One recent approach replaced end-to-end classification with a two-stage architecture comprising segmentation and classification, allowing the clinician to interrogate the segmentation map to understand the basis of the subsequent classification [24].

MD vs. Machine: Artificial intelligence in health care

Recent advances in artificial intelligence and machine learning are changing the way doctors practice medicine. Can medical data actually improve health care?

Michael Littman - Deep Learning and Artificial Intelligence Symposium

May 1, 2017 - In less than a decade, the field of “artificial intelligence” or “AI” has been jolted by the extraordinary and unexpected success of a set of techniques ...

12 Minutes AI - Everything you need to know - WorkTools #29 by Christoph Magnussen

Now with fixed image stabilisation... sorry! Subscribe here to my weekly vlogs and updates or check out my website .

Humans & AI: A Healthy Relationship? | Marcel Blattner | TEDxZurich

In the last decade we have witnessed a tremendous development: the rise of the machines. The world is drenched in algorithms and they actively influence our ...

Artificial Intelligence vs humans | Jim Hendler | TEDxBaltimore

Artificial Intelligence vs Humans - Jim disagrees with Stephen Hawking about the role Artificial Intelligence will play in our lives. Jim is an artificial intelligence ...

Krzysztof Geras: Advances in breast cancer screening with deep neural networks

Recent advances in deep learning for natural images has prompted a surge of interest in applying similar techniques to medical images. Most of the initial ...

Deep Learning in Medical Imaging - Ben Glocker #reworkDL

Machines capable of analysing and interpreting medical scans with super-human performance are within reach. Deep learning, in particular, has emerged as a ...

Best Laptop for Machine Learning

What kind of laptop should you get if you want to do machine learning? There are a lot of options out there and in this video i'll describe the components of an ...

Artificial intelligence now able to diagnose cancer as accurately as doctors can

Artificial intelligence can diagnose cancer as accurately as trained doctors, British scientists report. The breakthrough discovery is set to revolutionise healthcare, ...

Chao Han: Deep Learning vs. Conventional Machine Learning | PyData Miami 2019

Over the past few years, deep learning has given rise to a massive collection of ideas and techniques which are disruptive to conventional machine learning ...