Skip to main content

Mathematical Optimization for clinical DEcision Support and Training

Periodic Reporting for period 4 - MODEST (Mathematical Optimization for clinical DEcision Support and Training)

Reporting period: 2020-01-01 to 2020-06-30

Physicians need to make many important decisions per day. One clinical example is the scheduling and dosage of chemotherapy treatments. A second example is the discrimination of atrial fibrillation from atypical atrial flutter, based on ECG data. Such important and complex decisions are usually based on expert knowledge, accumulated throughout the life of a physician and shaped by subjective (and sometimes unconscious) experience. It is not readily transferable and may be unavailable in rural areas. At the same time, the available imaging, laboratory, and basic clinical data is abundant and waits to be used. This data is not yet systematically integrated and often single data-points are used to make therapy decisions.
More and more clinical decision making tasks will be modeled in terms of mathematical relations. We follow a systematic approach that supports and trains individual decision making. The developed ideas, mathematical models, and optimization algorithms will be generic and widely applicable in medicine and beyond, but also exploit specific structures, resulting in a patient- and circumstance-specific personalized medicine.

This allows, e.g. a physician to first simulate the impact of his decisions on a computer and to consider optimized solutions. In the future, it will be the rare and unwanted exception that an important decision can not be backed up by consultation of a model-driven decision support system or based upon a systematic model-driven training.

MODEST has a mathematical core. It builds on a comprehensive, interdisciplinary work program, based on disciplinary expertise in mixed-integer optimal control and existing collaborations with medical and educational experts. It is both timely, given the increasing availability of data and the maturity of mathematical methods, models, and software; as well as high-impact, due to the large number of clinical areas that may benefit from optimization-based decision support and training tools.
During the project, cooperations with clinical partners were established, data was acquisitoned and processed, and mathematical models and algorithms were developed. This relates to projects in oncology, where data concerning leucocyte counts during maintenance therapy of different kinds of leukemia and erythrocyte counts during treatment of Polycythemia vera were assessed, and to cardiology, where ECG data and 3d mapping data were analyzed. In addition to the cooperation that were already mentioned in the proposal further promising cooperations with clinical partners were started. Data has been analyzed,showing a great potential for optimization approaches. We did publish results in several high-ranking journals from different scientific fields and submitted further publications, covering the following areas:

* mathematical modeling of acute myeloid leukemia and comparison to clinical data
* mathematical modeling of acute lymohblastic leukemia and comparison to clinical data
* mathematical modeling of polycythemia vera and comparison to clinical data
* mathematical modeling and simulation of the Wolffe-Parkinson-White Syndrome
* mathematical modeling of extrasystoles and comparison to clinical data
* new algorithms for real-time modeling and optimization
* comparisons between model-driven and data-driven (machine learning) approaches for decision support

Optimization algorithms have been developed, implemented and applied to clinical decision making tasks. Results on optimal timing of measurements, sensitivities of white blood cell count dynamics with respect to timing and dosage of chemotherapies, optimal treatments for Acute Myeloid Leukemia in adults, simulated treatments for Acute Lymphoblastic Leukemia in children, optimal phlebotomy schedules for Polycythemia vera patiens, and on optimized measurement procedures during catheter ablation were obtained. For the application of Acute Myeloid Leukemia a novel dynamic stratification of patients was discovered. For the application of cardiac arrhythmia diagnostics the accuracy of our algorithms could be further increased by development of an innovative and interpretable combination of mathematical optimization and machine learning.
We could show for several prototypical clinical decision making tasks that there is a huge potential to improve health care quality by means of optimization-driven decision support for clinical doctors.

A survey of the ERC action was published in the newsletter of the Mathematical Optimisation Society, S. Sager, Optimization and Clinical Decision Support, Optima 104, pp. 1-8. This article may impact future research directions in mathematical optimization.