AI4BS
AI for Body and Soul

In recent years, artificial intelligence (AI) has revolutionized medical practices, enhancing both diagnosis and treatment. This keynote will delve into the intersection of AI and healthcare, focusing on two critical areas: radiology and liquid biopsy.

1. Radiology Advancements:
   • AI algorithms have significantly improved radiological image analysis. From detecting subtle anomalies in X-rays to identifying early signs of diseases in MRI scans, these tools augment the expertise of radiologists.
   • We’ll explore real-world examples where AI-driven image recognition has led to faster and more accurate diagnoses, ultimately improving patient outcomes.
2. Liquid Biopsy for Early Cancer Detection:
   • Liquid biopsy, a non-invasive technique, analyzes blood samples for cancer-related biomarkers. AI plays a pivotal role in interpreting these complex data sets.
   • By identifying circulating tumor cells, DNA, RNA, and proteins, liquid biopsy enables early cancer detection, personalized treatment, and monitoring of treatment response.
3. Convergent Science and Wicked Problems:
   • The future lies in interdisciplinary collaboration. Convergent science, where experts from diverse fields collaborate, can tackle wicked problems—complex, multifaceted challenges that defy traditional solutions.
   • We’ll discuss how AI, combined with insights from medicine, biology, ethics, and social sciences, can address pressing societal issues, from healthcare disparities to climate change.

Together we’ll explore the frontiers of science and medicine, and envision a healthier, more connected world.

Cortical visual prosthetic systems are investigated as a means of restoring vision in subjects with impaired eye-brain connection. A central hypothesis of such endeavor is that induction of an activity pattern in visual cortex similar to that evoked by a given visual stimulus, will elicit perception of similar visual stimulus. Existing V1 prostheses stimulation strategies take into account only the retinotopic aspect of visual coding [1,2], overlooking other important coding properties of V1 neurons such as selectivity for stimulus orientation. Here we introduce a novel stimulation protocol that stimulates cortical tissue according to both retinotopy and orientation columns. To obtain this protocol, we implemented a bottlenecked rotationally equivariant convolutional neural network, that learns to predict neural responses to arbitrary stimuli solely based on the neuron’s receptive field position and orientation preference. Our model outperforms the standard energy model of V1 complex cells. The high correlation between target and predicted responses suggests that position and orientation alone can explain a large portion of V1 neural response variability. To test whether stimulation of cortical tissue according to our model would indeed induce desired activity in the dynamic recurrent cortical networks, we utilized a previously published simulation framework [3] composed of a large-scale recurrent spiking V1 model to simulate optogenetic prosthetic stimulation delivered via an LED array placed on the cortical surface. Our simulations show that the newly proposed retinotopic-and-orientation-based stimulation protocol recruits neural patterns that more accurately mimic natural vision processing than analogous protocols only relying on coding of retinotopy.

1. Fernández, Eduardo, et al. “Visual percepts evoked with an intracortical 96-channel microelectrode array inserted in human occipital cortex.” The Journal of clinical investigation 131.23 (2021)
2. Chen, Xing, et al. “Shape perception via a high-channel-count neuroprosthesis in monkey visual cortex.” Science 370.6521 (2020): 1191-1196.
3. Antolik, Jan, et al. “Assessment of optogenetically-driven strategies for prosthetic restoration of cortical vision in large-scale neural simulation of V1.” Scientific reports 11.1 (2021): 10783.

Survival analysis is integral to understanding the timing of events of interest, often relying on models such as the Cox proportional hazard model. However, the rigidity of statistical assumptions inherent in these models poses limitations. In this study, we utilize a machine-learning approach and propose a novel approach aimed at circumventing these limitations by decomposing the time-to-event variable into distinct "time" and "event" components.
The decomposition enables us to view the time-to-event variable through a new lens, where the time component serves as a continuous covariate on the input of a classifier, and the event component, as a binary indicator, is on the classifier's output. This paradigm shift allows us to apply machine-learning classification algorithms that are inherently assumption-free, unlike traditional survival models. We employ a range of classification algorithms, including logistic regression, support vector machines, decision trees, random forests, and artificial neural networks, to model the event component and predict the occurrence of events of interest.
To validate our approach, we focus on predicting the decrease of IgG and IgM blood antibodies against COVID-19 (SARS-CoV-2), a pertinent clinical challenge. Using a dataset comprising 663 individuals, we train and test our models, including the Cox proportional hazard model, through rigorous cross-validation techniques.
Our findings demonstrate that the proposed machine-learning approach yields comparable or slightly superior predictive accuracy compared to the Cox model, particularly in predicting COVID-19 IgG antibody decrease. Multivariate logistic regression, support vector machines, random forests, and artificial neural networks consistently outperform the Cox model in this regard, considering predictive accuracy. However, predicting COVID-19 IgM antibody decrease proves challenging for both Cox regression and machine-learning algorithms, highlighting the complexity of antibody dynamics.
These results underscore the potential of our approach to improve event prediction in survival analysis, particularly in clinical settings. By minimizing reliance on strict statistical assumptions, our method offers a versatile and robust framework for predicting critical events such as antibody decreases, thereby facilitating proactive healthcare interventions. This advancement has implications beyond COVID-19, extending to various medical contexts where accurate event prediction is essential for effective patient management and resource allocation.

The rapid advancement of artificial intelligence (AI) and its integration into practical healthcare applications presents a formidable challenge for legislators and judicial systems, particularly in high-stakes domains like health care. While AI in various contexts can swiftly iterate and learn from mistakes, the delicate nature of health care demands that any additional risk is justified by a proportional benefit and errors are minimised.
At the same time, the relationship between a patient and their physician should be based on trust. It remains very uncertain what effect the AI systems will have on trust in society in general, including health care.
Nevertheless, use cases of AI in medicine typically pertain to, or are very similar to, areas that have already been vigorously regulated. In many instances, AI will be used in the form of medical devices (or as their feature or part), which are subjected to relatively detailed regulation. In other cases, AI will often be used by individual health professionals in a way similar to today’s use of the internet and other sources of information, which in itself is not a new situation for the law either.
Our exploration focuses on the pivotal question: what constitutes the optimal approach to AI regulation? Should lawmakers and courts meticulously define detailed legal solutions, or should they rather use broad legal frameworks grounded in general principles? Does the AI Act and other planned legal tools represent the step in the right direction?
The inquiry does not overlook critical aspects such as the questions of complexity and entropy within legal system and society, alongside pressing geopolitical issues. Our primary objective, however, is to delineate a regulatory path that safeguards and enhances patient rights and trust within the doctor-patient relationship amidst the integration of AI.

The domain of pharmaceutical science and drug discovery poses distinct opportunities and challenges for artificial intelligence. Distinctions include the limited data publicly available for many pharmaceutical endpoints and the types of explainability approaches relevant to pharmaceutical science. Here will discuss how these distinctions impacted the newly developed doctoral course teaching artificial intelligence at the Faculty of Pharmacy, Charles University. The syllabus, learning outcomes, prerequisites, and takeaways of what worked and what didn't during the first semester teaching this course will be reviewed.

Ve studii z roku 2023, která zahrnovala 1012 teenagerů z 19 měst České republiky, jsme prozkoumali, jak znalosti o umělé inteligenci u teenagerů ovlivňují emocionální vnímání postoj k této technologii.
Zjistili jsme, že vyšší sebevědomí ve znalostech o AI souvisí s pozitivnějšími emocemi jako zvídavost a úžas, což naznačuje větší ochotu k získávání nových informací. Analýza sentimentu navíc ukázala, že u teenagerů s pocitem hlubšího porozumění této technologii převažují pozitivní asociace.
Zajímavé však je, že toto sebevědomí nijak nekoreluje s obavami. Výsledky našeho výzkumu tedy nenaznačují, že by otevřená diskuse o rizicích a etických dilematech spojených s AI negativně ovlivňovala celkový postoj k této technologii. Naopak se ukazuje, že obavy mohou podporovat chuť získávat nové informace. Navíc ti, kteří se cítí jako nadprůměrně informovaní, hodnotí téma AI jako více důležité. A to jak pro jejich budoucí profesní dráhu, tak i pro společnost obecně.
Prezentace tohoto výzkumu by měla zdůraznit to, že je potřeba najít rovnováhu mezi podporou zvídavosti a zároveň upřímnou diskusí o možných výzvách a rizicích, které umělá inteligence přináší. Důležité je, že diskuse o rizicích a etických dilematech AI nejenže neodrazuje mladé lidi od technologie, ale může je naopak motivovat k hlubšímu zájmu a pochopení AI a jejího významu pro společnost.

The presentation focuses on how AI can be leveraged to create dynamic learning environments and simulate interactions with patients or clients. By incorporating generative AI technologies, educators can provide students with immersive and realistic experiences, better preparing them for real-world scenarios. The lecture will discuss the benefits, challenges, and ethical considerations associated with integrating generative AI into the curriculum of helping professions.