Causability and explainability of artificial intelligence in medicine.
Holzinger A, Langs G, Denk H, Zatloukal K, Müller H. · 2019
Read Causability and explainability of artificial intelligence in medicine. by Holzinger A, Langs G, Denk H, Zatloukal K, Müller H. (2019) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc7017860.txt.
Introduction
Holzinger and co-authors write a concept article on artificial intelligence in medicine. It appeared in Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery. Modern learning systems can match doctors on some tasks, yet their reasoning stays hidden. The authors argue that hidden reasoning is not enough for clinical use.
The article separates two ideas that are often merged. Explainability is a property of a system, they say. Causability is a property of a person, and it measures the quality of an explanation for that person. They give definitions for both terms. A worked case in tissue pathology shows the gap in practice. Their goal is a standard for explanations that doctors can truly follow and trust.
Catalog record
| CREATOR | Holzinger A, Langs G, Denk H, Zatloukal K, Müller H. |
|---|---|
| YEAR | 2019 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 63 KB |
| LICENSE | CC BY 4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Causability and explainability of artificial intelligence in medicine By Andreas Holzinger, Georg Langs, Helmut Denk, Kurt Zatloukal, Heimo Müller Published in: Wiley Interdisciplinary Reviews. Data Mining and Knowledge Discovery DOI: 10.1002/widm.1312 ABSTRACT Abstract Explainable artificial intelligence (AI) is attracting much interest in medicine. Technically, the problem of explainability is as old as AI itself and classic AI represented comprehensible retraceable approaches. However, their weakness was in dealing with uncertainties of the real world. Through the introduction of probabilistic learning, applications became increasingly successful, but increasingly opaque. Explainable AI deals with the implementation of transparency and traceability of statistical black‐box machine learning methods, particularly deep learning (DL). We argue that there is a need to go beyond explainable AI. To reach a level of explainable medicine we need causability. In the same way that usability encompasses measurements for the quality of use, causability encompasses measurements for the quality of explanations. In this article, we provide some necessary definitions to discriminate between explainability and causability as well as a use‐case of DL interpretation and of human explanation in histopathology. The main contribution of this article is the notion of causability, which is differentiated from explainability in that causability is a property of a person, while explainability is a property of a system This article is categorized under: - Fundamental Concepts of Data and Knowledge > Human Centricity and User Interaction Keywords: artificial intelligence, causability, explainability, explainable AI, histopathology, medicine INTRODUCTION AND MOTIVATION Artificial intelligence (AI) is perhaps the oldest field of computer science and very broad, dealing with all aspects of mimicking cognitive functions for real‐world problem solving and building systems that learn and think like people. Therefore, it is often called machine intelligence (Poole, Mackworth, & Goebel, 1998) to contrast it to human intelligence (Russell & Norvig, 2010). The field revolved around the intersection of cognitive science and computer science (Tenenbaum, Kemp, Griffiths, & Goodman, 2011). AI now raises enormous interest due to the practical successes in machine learning (ML). In AI there was always a strong linkage to explainability, and an early example is the Advice Taker proposed by McCarthy in 1958 as a “program with common sense” (McCarthy, 1960). It was probably the first time proposing common sense reasoning abilities as the key to AI. Recent research emphasizes more and more that AI systems should be able to build causal models of the world that support explanation and understanding, rather than merely solving pattern recognition problems (Lake, Ullman, Tenenbaum, & Gershman, 2017). ML is a very practical field of AI with the aim to develop software that can automatically learn from previous data to gain knowledge from experience and to gradually improve it's learning behavior to make predictions based on new data (Michalski, Carbonell, & Mitchell, 1984). The grand challenges are in sense‐making, in context understanding, and in decision making under uncertainty (Holzinger, 2017). ML can be seen as the workhorse of AI and the adoption of data intensive ML methods can meanwhile be found everywhere, throughout science, engineering and business, leading to more evidence‐based decision‐making (Jordan & Mitchell, 2015). The enormous progress in ML has been driven by the development of new statistical learning algorithms along with the availability of large data sets and low‐cost computation (Abadi et al., 2016). One nowadays extremely popular method is deep learning (DL). DL is a family of ML models based on deep convolutional neural networks having a long history (Schmidhuber, 2015). DL is very popular today because they are achieving amazing results even at human level performance (LeCun, Bengio, & Hinton, 2015). A best practice example is a recent work of the Thrun group, where they achieved with a DL approach performance on par with medical doctors, demonstrating that such approaches are able to classify skin cancer with a level of competence comparable to human dermatologists (Esteva et al., 2017). A further example is the promising results of identifying diabetic retinopathy and related eye diseases (Ting et al., 2017). All these are very good examples of the progress and usefulness of AI, but even the most prominent proponents of these (automatic) approaches recently emphasize that usable intelligence is difficult to reach because we need not only to learn from prior data, to extract knowledge, to generalize, and to fight the curse of dimensionality, but to disentangle the underlying explanatory factors of the data in order to understand the context in an application domain (Bengio, Courville, & Vincent, 2013), where to date a doctor‐in‐the‐loop is indispensable (Holzinger, 2016). Medicine as application domain is among the greatest challenges of AI/ML/DL. In medical decision support we are confronted with uncertainty, with probabilistic, unknown, incomplete, imbalanced, heterogeneous, noisy, dirty, erroneous, inaccurate and missing data sets in arbitrarily high‐dimensional spaces (Holzinger, Dehmer, & Jurisica, 2014), (Lee & Holzinger, 2016). Often we are simply lacking of large data sets (Holzinger, 2016). A grand goal of future medicine is in modeling the complexity of patients to tailor medical decisions, health practices and therapies to the individual patient (Holzinger, 2014). This poses challenges particularly in the integration, fusion and mapping of various distributed and heterogeneous data up to the visual analysis of these heterogeneous data (Turkay, Jeanquartier, Holzinger, & Hauser, 2014). Consequently, explainable‐AI in the context of medicine must take into account that diverse data may contribute to a relevant result. This requires that medical professionals must have a possibility to understand how and why a machine decision has been made (Holzinger, Biemann, Pattichis, & Kell, 2017). Explainability is at least as old as AI itself and rather a problem that has been caused by it. In the pioneering days of AI (Newell, Shaw, & Simon, 1958), reasoning methods were logical and symbolic. These approaches were successful, but only in a very limited domain space and with extremely limited practical applicability. A typical example is MYCIN (Shortliffe & Buchanan, 1975), which was an expert system developed in Lisp to identify bacteria causing severe infections and to recommend antibiotics. MYCIN was never used in clinical routine, maybe because of its stand‐alone character and the high effort in maintaining its knowledge base. However, these early AI systems reasoned by performing some form of logical inference on human readable symbols, and were able to provide a trace of their inference steps. This was the basis for explanation, and there is some early related work available, for example, (Johnson, 1994; Lacave & Diez, 2002; Swartout, Paris, & Moore, 1991). Here, we should mention that there are three types of explanations: (1) a peer‐to‐peer explanation as it is carried out among physicians during medical reporting; (2) an educational explanation as it is carried out between teachers and students; (3) A scientific explanation in the strict sense of science theory (Popper, 1935). We emphasize that in this article we mean the first type of explanation. In medicine there is growing demand for AI approaches, which are not only performing well, but are trustworthy, transparent, interpretable and explainable for a human expert; in medicine, for example, sentences of natural language (Hudec, Bednrov, & Holzinger, 2018). Methods and models are necessary to reenact the machine decision‐making process, to reproduce and to comprehend both the learning and knowledge extraction process. This is important, because for decision support it