Machine Learning and Artificial Intelligence provide new opportunities for investigating minimal tumour burden and therapy resistance by cancer patients
One advantage yet to focus on in scientific literature is the beneficial use of virtual autopsy (virtopsy) for investigating minimal tumour burden. Our hypothesis is that virtopsy assists in the understanding of therapy resistance of cancer patients or cause of death in patients with minimal tumour burden.1 The well-established textbook scenario describes a patient dying from cancer by the tumour mass compressing surrounding tissue (e.g. brain tumours), or destroying surrounding tissue resulting in organ failure (e.g. multi-metastatic diseases), or destroying blood vessels causing lethal bleeding. Furthermore, patients may die from tumor-induced cachexia, which is a consequence of the tumour’s interference with the body’s energy homeostasis. However, there are some instances when it is difficult to explain a tumour-related death, particularly when a patient dies of head or neck cancer.
For example, when major tumour mass is not detectable, there are no signs of cachexia or evidence of an immediate consequence of chemotherapy or radiotherapy. When patients might die exhibiting a small metastasis in the lung or a small local tumour, it is concluded that they died of cancer because there is no other apparent cause of death (e.g. cardiac disease). This is concluded despite none of the classical cancer-related causes of death being established. In view of this, therefore, is an unexplained mechanism of how a tumour can kill a patient. The majority of current cancer therapies are aimed at killing tumour cells. This is done either directly (by chemical agents or radiation), or indirectly (by depriving the tumour of nutrients, or activating and redirecting the immune response against cancer). This variety of therapeutic approaches is reflected in modern therapies such as PDL-1 blockers, VEGF-inhibitors, tumour-vaccines or Proteasome-inhibitors.
However, the observation that a fair number of patients with minimal detectable tumor mass die of cancer, given that treatment options described were considered, highlights gaps in knowledge that require filling as well as the development of new potential therapeutic approaches. As a first step, we propose epidemiological studies be undertaken – these are required in order to obtain quantifiable data on how many and which type of cancer patients die of cancer with minimal tumour mass. As autopsies are rarely performed on patients whose cancer has been well-characterized during the course of the disease, systematic data on this poorly-characterized cause of cancer-related death do not exist.
Furthermore, the assessment of total tumour mass in the body is difficult in disseminated diseases by traditional autopsy. In cases where minimal tumour burden has caused patients to die, we need to gain more knowledge (evidence-based practice). It is the combination of autopsy, pathology and virtopsy that truly defines or examines the entire body. When dealing with a localized disease, the traditional autopsy is appropriate in order to cut out, weigh and measure parts that are affected. This, however, cannot be performed when dealing with a disseminated disease with many small lesions in various organ systems.
Virtopsy can be a very effective intervention to quantify tumour mass. Imaging would provide very important baseline data to compare different patients’ tumour mass and to exclude other non-cancer-related causes of death.1 Machine learning methods will also
be of great assistance too, particularly for potential in-silico modelling.2,3 This would save time and effort; enabling what is not currently feasible in a wet-lab. Moreover, the principles of machine learning image analysis would enhance virtopsy.1 The enormous practical success of Machine Learning and Artificial Intelligence (AI) has led to more evidence-based decision-making in the medical domain.4 A very recent example with deep learning models demonstrated impressive results:5 the authors utilized a GoogleNet Inception v3 CNN architecture for the classification of skin lesions, using only pixels and disease labels as inputs. They pre-trained their network with 1.28 million images (1,000 object categories), and trained it on 129,450 clinical images, consisting of 2,032 different diseases.
The performance was tested against 21 board-certified dermatologists on biopsy-proven clinical images with two critical binary classification use cases: keratinocyte carcinomas versus benign seborrheic keratoses; and malignant melanomas versus benign nevi. The results demonstrated that such deep learning models can achieve a performance even beyond human experts. However, besides being resource-intensive and data-hungry, black-box machine learning and AI approaches have one enormous disadvantage in the medical domain – they are lacking transparency.
Even if we understand the mathematical theory of machine learning model it is complicated, yet impossible to get insight into the internal working of such a model. This leads to a major question – can we trust such results? Consequently, there is growing demand for interactive machine learning advances,6 which is not only well performing, but transparent, interpretable and trustworthy and include a human-in-the-loop.7 For the medical domain, it is necessary to re-enact the machine decision-making process, to reproduce and to comprehend the learning and knowledge extraction process.8 For medical decision support, it is of ultimate importance to understand the causality of learned representations.9,10 If human intelligence will be complemented by machine learning and at least in some cases even take precedence, humans must be able to understand and principally be able to interactively influence the machine decision process. This needs to make sense in order to close the gap between human thinking and machine “thinking”.11 Increasing legal and privacy characteristics are a massive motivation for this practice.
The new European General Data Protection Regulation (GDPR and ISO/IEC 27001) entering into force on May 25, 2018, will make black-box approaches difficult to use in any business, because they recognize that they are not able to explain why a decision has been made; this will make glass-box approaches essential12 and stimulate international research in interactive machine learning with the goal of making decisions interpretable, comprehensible and reproducible.
In our example, this is not only useful for machine learning research, and for clinical decision making but at the same time a big asset for the training of medical students. We, therefore, call for a collaborative effort to generate quantifiable data on tumour burden of patients who died of cancer without evidence of classical causes of cancer-related deaths. These data could lay the foundation for discovering novel mechanisms of how cancer may interfere with body function. There is increasing evidence from metabolomic studies that tumours may markedly interfere with metabolism as demonstrated by certain metabolic signatures that correlate with disease prognosis.13 Interestingly this interference was not related to signs of cachexia and therefore is most likely different from presently-described mechanisms of cancer-induced cachexia.14 A broadening of the scientific basis on the relevance of metabolic interference of cancer could widen the field of cancer therapies, which target the cancer interference with the patient’s metabolism.15
In summary, this article presents and supports an add-on example of how virtopsy can propel medicine into the future, its impact, implications and application in investigating minimal tumour burden and therapy resistance by cancer patients. Great advances will be made by taking advantage of current progress in Machine Learning and Artificial Intelligence, however, new approaches are needed that make use of a human-in-the-loop and above all in making transparent why and how a decision has been made.
1. O’Sullivan S, Holzinger A, Zatloukal K, Saldiva P, Sajid MI, Wichmann D. Machine learning enhanced virtual autopsy. Autopsy Case Rep. 2017;7(4):3-7. http://dx.doi.
2. Jeanquartier F, Jean-Quartier C, Cemernek D, Holzinger A. In silico modelling for tumour growth visualization. BMC Syst Biol. 2016;10(1):1-15. PMid:27503052. http://dx.doi.
3. Jeanquartier F, Jean-Quartier C, Kotlyar M, et al. 2016. Machine learning for in silico modeling of tumour growth. In: Holzinger A, editor. Machine learning for health informatics, Springer Lecture Notes in Artificial Intelligence LNAI 9605. Cham: Springer International Publishing. pp. 415-434. http://dx.doi.org/10.1007/978- 3-319-50478-0_21.
4. Jordan MI, Mitchell TM. Machine learning: trends, perspectives, and prospects. Science. 2015;349(6245):255- 60. PMid:26185243. http://dx.doi.org/10.1126/science.
5. Esteva A, Kuprel B, Novoa RA, et al. Dermatologist-level classification of skin cancer with deep neural networks. Nature. 2017;542(7639):115-8. PMid:28117445.http:// dx.doi.org/10.1038/nature21056.
6. Holzinger A. Interactive machine learning for health informatics: when do we need the human-in-the-loop? Brain Informatics. 2016;3(2):119-31. PMid:27747607. http://dx.doi.org/10.1007/s40708-016-0042-6.
7. Girardi D, Küng J, Kleiser R, et al. Interactive knowledge discovery with the doctor-in-the-loop: a practical example of cerebral aneurysms research. Brain Informatics.
2016;3(3):133-43. PMid:27747590. http://dx.doi. org/10.1007/s40708-016-0038-2.
8. Holzinger A, Plass M, Holzinger K, Crisan GC, Pintea C-M, Palade V. Towards interactive machine learning (iML): Applying ant colony algorithms to solve the travelling salesman problem with the human-in-the-loop approach. springer Lecture Notes in Computer Science LNCS 9817. Heidelberg: Springer; 2016. p. 81-95. http:// dx.doi.org/10.1007/978-3-319-45507-56.
9. Pearl J. 2009. Causality: models, reasoning, and inference. 2nd ed. Cambridge: Cambridge University Press.
10. Gershman SJ, Horvitz EJ, Tenenbaum JB. Computational rationality: a converging paradigm for intelligence in brains, minds, and machines. Science. 2015;349(6245):273-8. PMid:26185246. http://dx.doi.org/10.1126/science. aac6076.
11. Holzinger A. Trends in interactive knowledge discovery for personalized medicine: cognitive science meets machine learning. IEEE Intelligent Informatics Bulletin.
12. Holzinger A, Plass M, Holzinger K, Crisan GC, Pintea C-M, Palade V. A glass-box interactive machine learning approach for solving NP-hard problems with the human-in-the-loop. 2017. p. 1-26. [cited 2017 July 2017]. Available from: https ://arxiv.org/abs/1708.01104
13. Bertini I, Cacciatore S, Jensen BV, et al. Metabolomic NMR fingerprinting to identify and predict survival of patients with metastatic colorectal cancer. Cancer Res. 2012;72(1):356-64. PMid:22080567. http://dx.doi. org/10.1158/0008-5472.CAN-11-1543.
14. Loumaye A, Thissen JP. Biomarkers of cancer cachexia. Clin Biochem. 2017;50(18):1281-88. http://dx.doi. org/10.1016/j.clinbiochem.2017.07.011.
15. Argilés JM, López-Soriano FJ, Stemmler B, Busquets S. Novel targeted therapies for cancer cachexia. Biochem J. 2017;474(16):2663-78. PMid:28751550. http://dx.doi.
O’Sullivan S, Holzinger A, Wichmann D, Saldiva PHN, Sajid MI, Zatloukal K. Virtual autopsy: Machine Learning and Artificial Intelligence provide new opportunities for investigating minimal tumour burden and therapy resistance by cancer patients. Autopsy Case Rep [Internet]. 2018;8(1):e2018003. http://dx.doi.org/10.4322/acr.2018.003