Artificial intelligence in medicine

from Wikipedia, the free encyclopedia
X-ray of a hand with automatic determination of the skeletal maturity by software with the help of artificial intelligence

Artificial intelligence in medicine is a rapidly growing sub-area of artificial intelligence (AI) in which digital information is evaluated in order to make diagnoses that are as meaningful as possible and / or suggest optimized therapies . At the same time, there are concerns and fears that machines could replace or push back people.

Diagnosis

Computer vision for imaging diagnostics

Artificial intelligence plays a role in imaging diagnostics . The evaluation of images with statistical and learning methods is also summarized under the Radiomics department . Doctors are supported by decision support systems. By using these methods it is possible, for example, to determine the type of cancer cells more precisely, since the distinguishing features are often difficult to see with the human eye. This is important in precision medicine in order to suggest targeted therapy. Depending on the type of cancer, different therapies are sometimes necessary or useful. Radiomics is used, for example, to classify tumors in the lungs , breast , brain and skin, among other things .

oncology

The largest web-based and international study to date on automated skin cancer diagnosis under the direction of the Medical University of Vienna compared 511 doctors from 63 countries against 139 algorithms (mostly based on neural networks , CNN) in the detection of skin cancer on dermatoscopic images. In contrast to other studies, not only two types of skin changes (birthmarks and melanoma) had to be detected, but the seven most common pigmented skin changes. The study showed - in the experimental setting - not only a clear superiority of the best algorithms, but also that even “average” machines were able to recognize categories as well or better than medical professionals.

In an international study by the University of Heidelberg, in which 58 dermatologists from 17 countries competed against a folding neural network in a comparative test, the neural network was also found to be superior to dermatologists in the detection of melanoma , but not as soon as dermatologists received additional information such as age, Gender and location of the lesion were obtained. The true-negative rate among dermatologists, however, was 75.7 percent, well above that of the network, which only recognized 63.8 percent of harmless birthmarks as such. The researchers emphasized that artificial intelligence can help doctors diagnose skin cancer, but cannot replace it. There are other studies in larger or international settings, or small or local settings.

Although the opinion of these techniques also seems largely positive among dermatologists, and the first results are showing in favor of the collaboration between humans and machines, some scientists and clinicians urge caution when using these techniques. Many companies are also working on the commercialization of these projects, some based on hardware such as infrared laser beams, some based on dermatoscopy.For smartphones, there have been a number of often paid apps for years that are supposed to detect skin cancer on the basis of a photo, but mostly not based on new techniques such as neural networks. A scientific review could show that none of the tested applications showed sufficient accuracy and thus benefits for patients. The dermatologist and health care researcher Matthias Augustin also takes a critical view of this, as it could lead to application errors and incorrect diagnoses that laypeople would not be able to correctly assess.

In a scientific competition for the automated detection of breast cancer cells (diagnosis of metastases in sentinel node biopsies), 32 programs from 23 teams competed against a team of 11 pathologists who each had two hours to analyze 129 specimens. A comparison group consisted of an accomplished pathologist who was allowed to take as much time as he wanted, which, however, does not correspond to everyday clinical practice. Most of the programs used convolutional neural networks . Seven of the programs gave better results than the group of pathologists. The human pathologists often overlooked micrometastases, which rarely happened to the better programs. Five of the programs were even better than the accomplished pathologist who took 30 hours to analyze it. Even this accomplished pathologist missed a quarter of the micrometastases. In March 2018, scientists from Eötvös University in Budapest presented an AI that can detect breast cancer cells with the same accuracy - namely around 90 percent - during a mammography from the x-ray images of the female breast as an experienced human radiologist. The rate of false positive results was also the rate of the medical professionals. The evaluation of the X-ray images by radiologists is a monotonous, exhausting, lengthy and error-prone work.

The colonoscopy is considered to be the safest way to detect malignant tumors in the mast and colon at an early stage. Every year, 61,000 people in Germany develop colon cancer. During a colonoscopy, the doctor removes all suspicious growths, so-called polyps, regardless of whether the growth is benign or malignant. Whether it is a malignant tumor (so-called adenoma) can only be determined later in the laboratory. In autumn 2018, Japanese gastroenterologists used an AI in a clinical test, which was trained to detect malignant tumors in the intestine. The hit rate was 93 percent. Images from the intestine are magnified 500 times and transmitted to an AI, which can then recognize within a second whether the polyp is a benign or malignant tumor. The doctor then receives feedback via a tone or a message on the screen. The AI ​​should be trained further in order to improve the recognition rate even further. Then the AI ​​could go into routine operation.

neurology

Scientists at the University of California in San Francisco presented a pilot study with deep , artificial neural networks in the journal Radiology in the fall of 2018 , which could use brain scans to detect Alzheimer's disease on average six years before the final diagnosis. Alzheimer's is often only diagnosed by doctors when the first symptoms appear. Even experienced doctors find it difficult to recognize and correctly classify the small changes in the brain that occur in the early stages. This is why AI-based detection can make an important contribution to early detection and thus treatment. The network achieved a sensitivity of 100% with a true negative rate of 82%. Further investigations are to follow in order to verify the results.

Pulmonology

In 2020, the Vienna AI laboratory Deep Insight published the source code of an artificial neural network that was trained to use CT images of the lungs to classify whether the patient suffers from COVID-19 if the virus has already infected the lungs. The network differentiates between changes in the lungs caused by COVID-19, other pathological findings and normal status.

Automatic data analysis

Basically, cancer is as individual as the patient themselves. That is the reason why therapy helps one patient and not another. Here, AIs can compare the genetic analysis of patients with millions of data from other patient files, forms of treatment, and research papers within minutes and thus come to a very precise diagnosis, which is called precision medicine and would not be possible without the use of computers. This is not limited to just cancer diagnosis, but can also be used for heart attacks, diabetes, etc. It is important for this that the data is available in digital form (anonymized). Google, IBM, Microsoft, Amazon etc. offer platforms for uploading and providing such data.

For example, in August 2016 at the Medical Institute of the University of Tokyo, the computer program IBM Watson was able to correct a misdiagnosis made by doctors. The doctors diagnosed the patient with acute myeloid leukemia . The therapy was unsuccessful, which is why Watson was consulted. It took the AI ​​10 minutes to match the woman's DNA with 20 million cancer studies. Watson recognized a very rare form of leukemia that has only affected 41 patients and is curable. However, IBM Watson's treatment suggestions can also be incorrect, for example if insufficient training data is available. Corresponding reports on incorrect recommendations, the use of which could endanger patients, were published by a medical specialist portal in 2018. According to IBM, the malfunction should have been fixed in a later version.

In January 2018, scientists from Stanford University presented an AI that can calculate with a 90 percent probability of terminally ill patients from the medical data whether they will die within the next 3 to 12 months. This could help terminally ill patients to spend the last few months with dignity and without aggressive treatment methods and possibly at home with palliative care .

The Apple Watch wristwatch records, among other things, a person's heart rate. Apple announced that there is an 85 percent probability that AIs can detect diabetes mellitus in the person wearing the watch from a heart rate analysis . The idea is based on the Framingham Heart Study, which recognized as early as 2015 that one can diagnose diabetes with the help of heart rate alone. Apple had previously succeeded in detecting an abnormal heart rhythm from the heart rate with a 97 percent probability, sleep apnea with 90 percent, and hypertension (high blood pressure) with 82 percent.

Language processing

In January 2018, researchers from the Mount Sinai School of Medicine demonstrated how psychological conversation logs with adolescents can tell whether they will develop psychosis in the next two years . The natural language processing helped to achieve on standardized tests up to 83 percent accuracy, as based disorganized thought processes, cumbersome formulations unclear associations or reduced speech complexity. The subtle differences can be seen after training with many such conversations.

In September 2018, researchers at MIT presented an AI that can diagnose depression in patients using spoken text or written text. Doctors and psychologists ask the patient questions about lifestyle, behavior and sensitivities in order to diagnose depression from the answers. After training with such interviews, the program also recognized depression based on everyday conversations with a hit rate of 83 percent - and when classifying the severity of the depression on a scale from 0 to 27 with a 71 percent hit rate. The AI ​​could support doctors with or permanently monitor users as an app to alert them in an emergency. The researchers also want to recognize dementia from language in the future.

According to the manufacturer, the health app Babylon Health should be able to use a voice system (chatbot) based on AI to create a diagnosis in conversation with patients that is about ten times more accurate than diagnoses by a general practitioner . The development of the app was also co-funded by the UK health system. The aim was to reduce costs. Although the app is intended to significantly reduce doctor visits, patients quickly found out how to use the app to get doctor appointments faster due to incorrect symptom descriptions.

The Ada app from the Berlin company Ada Health supports diagnoses based on the symptom description with an AI. According to the manufacturer, this should match the quality of well-trained western doctors. The Ada app sent unauthorized marketing companies such as Amplitude and Adjust, headquartered in San Francisco (USA), as well as Facebook.com regularly during the use of the app personal data, even if you do not have a Facebook account. The app was recognized by MIT and funded by the Bill & Melinda Gates Foundation . In 2019 Ada Health announced a cooperation with Sutter Health . Especially in developing countries, where there is a shortage of medical staff, the app can help to build a health system.

Criticism and disputes

It is controversial whether the high accuracy of the AI ​​for diagnosing diseases, for example, which has been indicated in some studies, is valid in practice. As a rule, the values ​​refer to previously determined, sometimes unrepresentative, historical data sets. For example, a study by Google's subsidiary DeepMind on the automated prediction of kidney failure , which was carried out on a data set of which only 6% came from female patients, is criticized. The lack of variation in the data sets could lead to computer programs that are very limited in their generalization and that do not provide the desired accuracy in real application scenarios. Furthermore, the insufficiently available explainability of the systems is cited as a weak point.

Pharmaceutical research

In pharmaceutical research , automated high-throughput screening has established itself as a method for finding so-called hits and thus candidates for lead structures . British researchers at Cambridge University developed automation further. The research robot "Eve", which was featured in the Journal of the Royal Society Interface in 2015 , uses statistical models and machine learning to produce and test assumptions, check observations, carry out experiments, interpret results, change hypotheses and repeat this over and over again. This allows the robot to predict promising substances and thus make finding lead structures more efficient. With the help of this robot, the researchers found out in 2018 that triclosan , which is also used in toothpaste, could fight malaria infections in two critical stages, namely infestation of the liver and blood. With the discovery by AI, a new drug could now be developed.

Web links

Individual evidence

  1. ^ P. Lombardo, I. Boehm, K. Nairz: RadioComics - Santa Claus and the future of radiology . In: European Journal of Radiology . 122, No. 1, 2020, p. 108771. doi : 10.1016 / j.ejrad.2019.108771 . PMID 31835078 .
  2. ^ Alanna Vial, David Stirling, Matthew Field, Montserrat Ros, Christian Ritz: The role of deep learning and radiomic feature extraction in cancer-specific predictive modeling: a review . In: Translational Cancer Research . tape 7 , no. 3 , July 6, 2018, ISSN  2219-6803 , p. 803–816 , doi : 10.21037 / 21823 ( amegroups.com [accessed December 9, 2018]).
  3. P Tschandl, N Codella, BN Akay, G Argenziano, RP Braun, H Cabo, D Gutman, A Halpern, B Helba, R Hofmann-Wellenhof, A Lallas, J Lapins, C Longo, J Malvehy, MA Marchetti, A Marghoob , S Menzies, A Oakley, J Paoli, S Puig, C Rinner, C Rosendahl, A Scope, C Sinz, HP Soyer, L Thomas, I Zalaudek, H Kittler: Comparison of the accuracy of human readers versus machine-learning algorithms for pigmented skin lesion classification: an open, web-based, international, diagnostic study. . In: The Lancet. Oncology . 20, No. 7, July 2019, pp. 938-947. doi : 10.1016 / S1470-2045 (19) 30333-X . PMID 31201137 .
  4. P Tschandl, C Rosendahl, H Kittler: The HAM10000 dataset, a large collection of multi-source dermatoscopic images of pigmented skin lesions common. . In: Scientific data . 5, August 14, 2018, p. 180161. doi : 10.1038 / sdata.2018.161 . PMID 30106392 .
  5. HA Haenssle, C Fink, R Schneiderbauer, F Toberer, T Buhl, A Blum, A Kalloo, ABH Hassen, L Thomas, A Enk, L Uhlmann, Groups. Reader study level-I and level-II: Man against machine: diagnostic performance of a deep learning convolutional neural network for dermoscopic melanoma recognition in comparison to 58 dermatologists. . In: Annals of oncology: official journal of the European Society for Medical Oncology . 29, No. 8, August 1, 2018, pp. 1836–1842. doi : 10.1093 / annonc / mdy166 . PMID 29846502 .
  6. P Tschandl, C Rosendahl, BN Akay, G Argenziano, A Blum, RP Braun, H Cabo, JY Gourhant, J Kreusch, A Lallas, J Lapins, A Marghoob, S Menzies, NM Neuber, J Paoli, HS Rabinovitz, C Rinner, A Scope, HP Soyer, C Sinz, L Thomas, I Zalaudek, H Kittler: Expert-Level Diagnosis of Nonpigmented Skin Cancer by Combined Convolutional Neural Networks. . In: JAMA dermatology . 155, No. 1, January 1, 2019, pp. 58-65. doi : 10.1001 / jamadermatol.2018.4378 . PMID 30484822 .
  7. MA Marchetti, NCF Codella, SW Dusza, DA Gutman, B Helba, A Kalloo, N Mishra, C Carrera, ME Celebi, JL DeFazio, N Jaimes, AA Marghoob, E Quigley, A Scope, O Yélamos, AC Halpern, Collaboration . International Skin Imaging: Results of the 2016 International Skin Imaging Collaboration International Symposium on Biomedical Imaging challenge: Comparison of the accuracy of computer algorithms to dermatologists for the diagnosis of melanoma from dermoscopic images. . In: Journal of the American Academy of Dermatology . 78, No. 2, February 2018, pp. 270–277.e1. doi : 10.1016 / j.jaad.2017.08.016 . PMID 28969863 .
  8. HA Haenssle, C Fink, F Toberer, J Winkler, W Stolz, T Deinlein, R Hofmann-Wellenhof, A Lallas, S Emmert, T Buhl, M Zutt, A Blum, MS Abassi, L Thomas, I Tromme, P Tschandl , A Enk, A Rosenberger, Groups. Reader Study Level I and Level II: Man against machine reloaded: performance of a market-approved convolutional neural network in classifying a broad spectrum of skin lesions in comparison with 96 dermatologists working under less artificial conditions. . In: Annals of oncology: official journal of the European Society for Medical Oncology . 31, No. 1, January 2020, pp. 137-143. doi : 10.1016 / j.annonc.2019.10.013 . PMID 31912788 .
  9. P Tschandl, H Kittler, G Argenziano: A Pretrained neural network shows similar diagnostic accuracy to medical students in categorizing dermatoscopic images after Comparable training conditions. . In: The British journal of dermatology . 177, No. 3, September 2017, pp. 867–869. doi : 10.1111 / bjd.15695 . PMID 28569993 .
  10. SS Han, MS Kim, W Lim, GH Park, I Park, SE Chang: Classification of the Clinical Images for Benign and Malignant Cutaneous Tumors Using a Deep Learning Algorithm. . In: The Journal of investigative dermatology . 138, No. 7, July 2018, pp. 1529–1538. doi : 10.1016 / j.jid.2018.01.028 . PMID 29428356 .
  11. C Yu, S Yang, W Kim, J Jung, KY Chung, SW Lee, B Oh: Acral melanoma detection using a convolutional neural network for dermoscopy images. . In: PloS one . 13, No. 3, 2018, p. E0193321. doi : 10.1371 / journal.pone.0193321 . PMID 29513718 .
  12. A Esteva, B Kuprel, RA Novoa, J Ko, SM Swetter, HM Blau, S Thrun: Dermatologist-level classification of skin cancer with deep neural networks. . In: Nature . 542, no. 7639, February 2, 2017, pp. 115–118. doi : 10.1038 / nature21056 . PMID 28117445 .
  13. S Polesie, M Gillstedt, H Kittler, A Lallas, P Tschandl, I Zalaudek, J Paoli: Attitudes towards artificial intelligence within dermatology: an international online survey. . In: The British journal of dermatology . January 17, 2020. doi : 10.1111 / bjd.18875 . PMID 31953854 .
  14. SS Han, I Park, W Lim, MS Kim, GH Park, JB Chae, CH Huh, SE Chang, JI Na: Augment Intelligence Dermatology: Deep Neural Networks Empower Medical Professionals in Diagnosing Skin Cancer and Predicting Treatment Options for 134 Skin Disorders . . In: The Journal of investigative dermatology . February 5, 2020. doi : 10.1016 / j.jid.2020.01.019 . PMID 32243882 .
  15. ^ A Lallas, G Argenziano: Artificial intelligence and melanoma diagnosis: ignoring human nature may lead to false predictions. . In: Dermatology practical & conceptual . 8, No. 4, October 2018, pp. 249-251. doi : 10.5826 / dpc.0804a01 . PMID 30479851 .
  16. ^ C Navarrete-Dechent, SW Dusza, K Liopyris, AA Marghoob, AC Halpern, MA Marchetti: Automated Dermatological Diagnosis: Hype or Reality? . In: The Journal of investigative dermatology . 138, No. 10, October 2018, pp. 2277–2279. doi : 10.1016 / j.jid.2018.04.040 . PMID 29864435 .
  17. A Berlin startup recognizes skin cancer better than a doctor gruenderszene.de of February 13, 2018
  18. DermEngine AI
  19. FotoFinder MoleAnalyzer
  20. A Ngoo, A Finnane, E McMeniman, JM Tan, M Janda, HP Soyer: Efficacy of smartphone applications in high-risk pigmented lesions. . In: The Australasian journal of dermatology . 59, No. 3, August 2018, pp. E175-e182. doi : 10.1111 / ajd.12599 . PMID 28240347 .
  21. NDR: How useful are "skin cancer apps"? ( ndr.de [accessed December 1, 2018]).
  22. Doctor versus computer: who is best at recognizing breast cancer metastases? aerzteblatt.de from December 13, 2017
  23. AI detects breast cancer: self-learning system detects tumors in mammograms as well as a radiologist. March 16, 2018, accessed July 2, 2019 .
  24. DETECTING INTESTINAL CANCER PRE-STAGES WITH ARTIFICIAL INTELLIGENCE internisten-im-netz.de of October 8, 2018
  25. ^ Yiming Ding, Jae Ho Sohn, Michael G. Kawczynski, Hari Trivedi, Roy Harnish: A Deep Learning Model to Predict a Diagnosis of Alzheimer Disease by Using 18F-FDG PET of the Brain . In: Radiology . November 6, 2018, ISSN  0033-8419 , p. 180958 , doi : 10.1148 / radiol.2018180958 ( rsna.org [accessed December 1, 2018]).
  26. Artificial intelligence recognizes Alzheimer's earlier than doctors . In: BR24 . ( br.de [accessed on December 1, 2018]).
  27. heise online: Artificial intelligence diagnoses COVID-19 using CT. Retrieved April 16, 2020 .
  28. ^ FAZ: Diagnoses by Watson. June 10, 2017, accessed July 2, 2019 .
  29. AI saves the patient's life by correcting the doctors' misdiagnosis vice.com of August 10, 2016
  30. AI against cancer: Criticism of IBM Watson for "dangerous recommendations" - derStandard.at. Retrieved December 9, 2018 .
  31. Artificial intelligence is supposed to predict the time of death of patients. Retrieved July 2, 2019 .
  32. Study: Apple Watch detects diabetes with 85 percent accuracy. February 7, 2018, accessed July 2, 2019 .
  33. AI can detect psychoses in young people at an early stage futurezone.de from January 23, 2018
  34. Cheryl M. Corcoran, Facundo Carrillo, Diego Fernández-Slezak, Gillinder Bedi, Casimir Klim: Prediction of psychosis across protocols and risk cohorts using automated language analysis . In: World Psychiatry . tape 17 , no. 1 , January 19, 2018, ISSN  1723-8617 , p. 67-75 , doi : 10.1002 / wps.20491 , PMID 29352548 ( wiley.com [accessed September 17, 2018]).
  35. Artificial intelligence is supposed to recognize depression by the way you speak. September 6, 2018, accessed July 2, 2019 .
  36. This app should make the doctor's visit superfluous - but the users did the opposite. December 17, 2017, accessed July 2, 2019 .
  37. Massive data protection deficiencies in the health app Ada. October 10, 2019, accessed October 10, 2019 .
  38. "We give medical advice to millions of people". October 10, 2018, accessed July 2, 2019 .
  39. ^ A question about the eHealth Europe congress in Freiburg. Retrieved July 2, 2019 .
  40. Michael Moorstedt: The blind spots of the artificial intelligences . In: sueddeutsche.de . August 11, 2019, ISSN  0174-4917 ( sueddeutsche.de [accessed on August 13, 2019]).
  41. Kevin Williams, Elizabeth Bilsland, Andrew Sparkes, Wayne Aubrey, Michael Young: Cheaper faster drug development validated by the repositioning of drugs against neglected tropical diseases . In: Journal of The Royal Society Interface . tape 12 , no. 104 , March 6, 2015, ISSN  1742-5689 , p. 20141289 , doi : 10.1098 / rsif.2014.1289 , PMID 25652463 ( royalsocietypublishing.org [accessed December 1, 2018]).
  42. Artificially intelligent robot scientist 'Eve' could boost search for new drugs . ( phys.org [accessed December 1, 2018]).
  43. AI robot finds ingredient in toothpaste may help fight malaria . In: The Independent . ( independent.co.uk [accessed December 1, 2018]).