Reinforcement learning, a type of machine learning where an algorithm learns to make sequences of decisions by maximizing cumulative rewards, has growing applications in treatment planning (52). Data scientists need to account for the differences between the oncology environment and a classic RL environment following the Markov assumption that the future state of a system depends only on its current state (53). They may do this by suggesting an adaptation to a RL model. In oncology, RL is used to develop dynamic treatment strategies that adapt based on the patient's response to ongoing treatments. RL algorithms can continuously adjust dosages of radiation or chemotherapy to minimize toxicity while maximizing effectiveness (54). Moreover, Tempo, a novel framework based on RL for personalized screening, has proven its effectiveness in the context of breast cancer. The Tempo policy, combined with an AI risk model, outperforms current practices in early detection and can be adapted to different screening preferences. It also improves detection while reducing overscreening (55). A personalized treatment approach allows clinicians to adjust treatment plans in real time, providing adaptive strategies as new patient data emerges. Similarly, personalized screening tailors assessments to individual risk profiles, enhancing early detection and treatment precision.

Data scientists also use Bayesian networks, probabilistic graphical models that represent a set of variables and their conditional dependencies. These networks help in predicting outcomes and understanding the likelihood of various treatment responses based on observed data. In a tumor board setting, Bayesian networks can integrate information from different sources—clinical data, molecular biomarkers, and patient history—to estimate the probabilities of different clinical outcomes and provide uncertainty quantifications, helping doctors make more informed decisions in cases where there is ambiguity (56). For example, in the study of Huehn et al. (57), they developed a digital patient model using Bayesian networks to integrate patient data and relevant probabilities for treatment decisions in head and neck squamous cell carcinoma (HNSCC). Validation showed the model effectively guides immunotherapy decisions, with 84% accuracy and significant concordance (Cohen's κ = 0.505, p = 0.009) when compared to actual treatment decisions for 25 patients. This model was created to represent a physician's decision-making in tumor boards, integrating clinical and molecular data to identify the best treatment for each patient.

Simulation-based approaches enable data scientists to create virtual treatment scenarios and evaluate potential outcomes. By simulating multiple strategies, clinicians can explore the consequences of various plans before applying them to patients. These simulations are especially valuable in complex cases with multiple options, helping model long-term effects of treatments like chemotherapy combined with immunotherapy or predict patient responses based on genetic profiles (58). Just like in the study of Federov et al. (59), they developed a novel simulation-based method for optimizing Tumor Treating Fields (TTFields) treatment planning for brain tumors. TTFields, delivered through transducer arrays on the skin, inhibit tumor growth, with their distribution varying based on array placement, patient anatomy, and tumor characteristics. Incorporating such algorithms into advanced treatment planning systems is expected to enhance physician management of TTFields patients, ultimately improving patient outcomes.

In addition to simulating treatment outcomes, data scientists also use dose optimization models to personalize the amount and timing of therapies such as chemotherapy or radiation. These models aim to balance efficacy with minimizing side effects, especially in treatments that involve toxic agents. By adjusting dosing schedules based on individual patient metabolism and tumor characteristics, clinicians can optimize therapy duration and frequency to increase the probability of success without compromising patient quality of life (60). The emergence of new anti-tumor treatments further complicates this, creating an urgent need for optimized dose-schedule strategies. Data scientists can be useful by using Bayesian adaptive designs which can offer an efficient approach to assess multiple doses and schedules concurrently within a single clinical trial (61). Moreover, recent advancements in deep learning (DL) have led to the development of DL-based dose prediction models. Unlike traditional methods, DL automatically extracts features from patient CT, MRI, or PET scans to map optimal dose values, guiding treatment planning systems for final dose distribution. It can also predict dose distributions based on anatomical data and dose prescriptions (62).

Multi-omics integration is another important facet of treatment planning. Data scientists combine data from genomics, transcriptomics, proteomics, and metabolomics to create a comprehensive molecular profile of a patient's tumor. By integrating these diverse datasets, data scientists help tumor boards develop more targeted therapies. For example, a multi-omics approach might reveal not just a single genetic mutation driving the cancer but also how it interacts with other molecular pathways, leading to more tailored treatments, such as targeting metabolic pathways in addition to the genomic aberration (63). Multi-omics offers a comprehensive view of cancer biology, but the large data volumes pose analytical challenges. AI helps address these by identifying patterns and extracting insights from complex omics data, advancing cancer research (64). Similar to the study by Cai et al. (65), which explored machine learning methods for integrating multi-omics data in cancer research, their focus was on both general-purpose and task-specific approaches. They benchmarked five methods using data from the Cancer Cell Line Encyclopedia, assessing accuracy in cancer classification, drug response prediction, and runtime efficiency. Their paper provides recommendations for selecting appropriate methods and encourages the development of new tools to advance drug discovery and personalized treatments.

Radiomics involves extracting quantitative features from medical imaging (e.g., CT, MRI, PET scans) that can reveal tumor heterogeneity and predict patient outcomes. Similarly, pathomics analyzes histopathological images to identify patterns that may not be discernible by a pathologist alone. Data scientists combine these image-derived features with clinical and genomic data to improve treatment predictions and select optimal therapeutic strategies. Like in the study of Tan et al. (66), they developed radiopathomics models, employing ML algorithms, that can highly classify stage I, II, and III gastric cancer. Other researchers also developed radiopathomics models to predict prognosis in patients with colorectal and lung cancers, as well as their overall survival (67, 68). Radiomics and pathomics are particularly valuable in situations where genetic testing may not be available, offering non-invasive options for assessing tumor characteristics (69, 70).

Spatial biology technologies, such as GeoMx® (NanoString Technologies®)¹, CosMx™ (NanoString Technologies®)², Visium® (10x Genomics®)³, and Xenium™ (10x Genomics®)⁴, are revolutionizing cancer research by integrating molecular profiling with spatial context. These approaches enable the mapping of tumor heterogeneity, microenvironment, and cell-cell interactions, providing insights that traditional bulk sequencing cannot offer. For instance, spatial transcriptomics (spTx) platforms like CosMx™⁵ and Xenium™⁶ enhance tumor analysis by combining high-resolution imaging with RNA profiling, capturing cellular organization and biomarker localization within tissue samples (71). Visium® HD⁷ offers sub-cellular resolution, enabling more detailed reconstruction of cell morphology and expression patterns in tumors (72).

One of the most notable applications of spatial biology in the field of oncology is the use of spTx in understanding breast cancer. Despite the improved patient outcomes in precision medicine, breast cancer treatment still faces challenges due to intratumoral heterogeneity (ITH), where different parts of the tumour respond differently to drugs. Using spTx, recent research shows that drug sensitivity varies across the tumor, from the core to the periphery. It also finds that genetically identical tumor cells can respond differently depending on their location (73). Thus, it is crucial to consider not only the genetic profile of the tumor but also its spatial organization and the surrounding microenvironment. This approach could lead to more personalized and effective treatments by addressing the tumor's complexity, improving outcomes, and reducing the chances of treatment failure.

Causal modeling focuses on determining cause-and-effect relationships, going beyond correlation to uncover how specific interventions lead to observed outcomes. This is particularly critical in oncology, where data scientists use advanced techniques like the Peter-Clark (PC) algorithm (74) and latent Gaussian models to identify causal pathways in treatment effects and disease progression (65). While frameworks like the SHapley Additive exPlanations (SHAP) and Local Interpretable Model Agnostic Explanation (LIME) help interpret machine learning models, they primarily explain correlations rather than causation (75).

For example, a large language model (LLM) was used to improve causal discovery of the factors impacting survival in Non Small Cell Lung Cancer (NSCLC), revealing potentially unexpected causal relationships such as smoking status having a causal effect on treatment choice (76). The role of data scientists extends further, applying causal inference methods, such as inverse probability weighting and structural failure time models, to address biases like treatment switching in survival analysis (77). By integrating these causal approaches into clinical workflows, data scientists enable tumor boards to base decisions on robust cause-and-effect insights, driving better patient outcomes in personalized oncology care.

Another application where data scientists play a key role is in developing adaptive treatment strategies. These strategies are designed to evolve in response to the patient's changing condition. For instance, if a patient's tumor becomes resistant to chemotherapy, data scientists can employ adaptive models to recommend alternative treatments, such as switching to targeted therapy or immunotherapy. By continuously analyzing patient data throughout the treatment course, adaptive models ensure that treatment plans remain flexible and responsive to real-time changes in disease progression (78). For example, adaptive radiotherapy (ART) is used as a cancer treatment method that adjusts the radiation plan during treatment to account for changes in the tumor's size, shape, and position, as well as nearby healthy tissues. This helps make the treatment more effective and safer by targeting the tumor better while reducing harm to healthy tissues, thus improving both the treatment's efficacy and safety (79).

Data scientists also develop decision support systems (DSS) that aid oncologists in selecting treatment pathways based on the integration of patient data, clinical guidelines, and outcomes from past treatments. These systems use a combination of machine learning algorithms and clinical knowledge to provide recommendations that consider both standard care protocols and personalized factors. In particular, data scientists design systems capable of providing real-time decision support by processing data from sensors, wearables, or EHR systems. For example, a clinical DSS could alert the physician in real time about the patient's deteriorating condition, helping to prevent adverse effects or complications (80). Monitoring biometric data in cancer patients offers crucial insights for oncologists. Initial metrics such as steps and heart rate aid in prognosis and treatment choices, while ongoing tracking helps identify potential adverse events and supports rehabilitation monitoring (81). Additionally, a recent thesis from the University of Toronto found that Machine Learning, especially random forests (RF), could be used to create a better clinical DSS (CDSS) for metastatic lung cancer than current hospital decision support framework (82). A one-step prediction of non-local therapy, surgery, or stereotactic body radiation therapy using RF achieved an Area under the Receiver Operator Curve of 0.857 and a mean accuracy of 72.55% against the tumor board's final decision as truth.

Data scientists can also help optimize combination therapies, where multiple treatments such as chemotherapy, radiotherapy, and immunotherapy are used in tandem. By modeling the interaction effects between various treatments, they can predict the synergistic or antagonistic outcomes of combining therapies. This is particularly important in oncology, where combination treatments are often necessary but must be carefully balanced to avoid excessive toxicity while maximizing therapeutic benefit (83). For instance, nanoplatforms are used to improve treatment precision and reduce toxicity by enhancing drug delivery, especially in combination with immunotherapy, chemotherapy, radiotherapy, and other cancer treatment procedures (84). With the help of data scientists, AI-powered predictive modeling can be utilized to help simulate intricate nanoscale processes, assisting in the design and refinement of nanomaterials and devices. This technique will enhance the development of nanoplatforms, optimizing their performance in drug delivery and cancer therapy while reducing associated risks.

These models, including GPT and BERT (85, 86), excel at processing large amounts of textual data from patient records and clinical notes. Their ability to handle unstructured data and perform complex tasks like summarization and sentiment analysis offers significant advantages over traditional statistical methods, which often struggle with the high dimensionality and noise inherent in medical data (91). In tumor boards, transformer-based models can summarize complex patient histories and predict treatment responses, providing oncologists with actionable insights.

Memory-Augmented Neural Networks (MANNs) are designed to enhance long-term dependencies, allowing for the retention of information over extended periods (92). This capability is particularly valuable for tracking the progression of diseases like cancer, where long-term data about tumor growth and treatment responses can inform future care strategies. MANNs are essential in long-term treatment planning, helping clinicians assess how a patient's condition might evolve over time (93).

Graph Neural Networks (GNNs) play a crucial role in understanding relationships between biological factors such as gene expressions, protein-protein interactions, and TMEs (94). These models analyze complex biological networks to provide insights into molecular mechanisms that drive cancer progression. By mapping interactions within biological pathways, GNNs help identify potential therapeutic targets, making them valuable for personalized treatment strategies in oncology (95).

Neural Ordinary Differential Equations (ODEs) offer a powerful approach to modeling continuous biological processes (96). These models are adept at simulating how a patient's condition changes over time, making them useful for predicting treatment responses and disease progression. Neural ODEs are particularly valuable in dynamic environments like oncology, where patient conditions can rapidly change due to treatment or disease relapse (97).

The Hyena Hierarchy model is a relatively new advancement in sequence modeling, designed to outperform traditional transformer models in processing long-range dependencies without the quadratic scaling issues inherent in transformers (98). In tumor boards, Hyena Hierarchy can be applied to analyze extensive patient datasets, such as multi-modal data from clinical trials or longitudinal studies. This model's ability to handle long sequences efficiently makes it ideal for extracting patterns from extensive patient histories, genetic sequences, and radiological imaging, leading to more personalized and precise treatment plans.

FNet models replace the self-attention mechanism found in transformers with Fourier Transforms, which are computationally efficient and scalable. FNet models are particularly useful in situations where quick and efficient data processing is required, such as analyzing large datasets of patient vitals or sensor data from wearable devices (99). In tumor boards, FNet could be used to extract meaningful insights from continuous data streams like monitoring patient vitals in real time, offering the potential for more dynamic, real-time decision-making during cancer treatment.

Retentive Networks are designed to process long-range data while retaining context across large sequences, much like transformer models but with more efficient memory use (100). This model can be useful in oncology, where a continuous stream of patient data is analyzed over the long term. Retentive Networks can help tumor boards track the effectiveness of treatments over time, analyzing data from repeated scans, lab results, and genomic updates to ensure that treatments remain optimal as the disease evolves.

Adaptive Computation Time is a strategy in DL models that improves computation efficiency and adaptability (101). It is particularly relevant in scenarios where different parts of a dataset require varying amounts of computational attention. In tumor boards, these models can dynamically allocate resources to the most critical patient data, such as prioritizing genomic markers or radiology results in patients with advanced cancers. This ensures that the model focuses on the most relevant data points to assist with urgent clinical decisions.

Variational Autoencoders (VAEs) play an important role in modeling high-dimensional data, which is common in oncology due to the complexity of genetic and clinical information (102). VAEs can be used to compress these large datasets while retaining the most relevant features for tumor classification or risk stratification. In tumor boards, VAEs assist by reducing data complexity, enabling clinicians to focus on the most critical elements for personalized treatment planning (103).

Here is a brief breakdown of supercomputer power distribution in the Global South by region. Computing power is often quantified by “flops”, where one flop indicates the ability to perform one floating point operation per second. For reference, the values below can be compared to the most powerful supercomputers in the world such as El Capitan⁸ in the United States with 1,742 petaflops (1.74e + 18 flops) based on the Top500 supercomputer list in 2024. Unfortunately, there is not a quality report comparing AI task processing speeds which can differ significantly from traditional metrics. For example, the Aurora system⁹ which ranks number three on the Top500 list may be the world's fastest computer for AI tasks. And those computers which could be used for AI tasks may be used for predictive modeling of phenomena such as weather instead. Because of this, some of the descriptions below may over- or under-estimate the systems' usefulness for integrating AI in tumor boards. Also of note, it is believed that China does not publish complete supercomputer specifications for all its systems to public sources.

South America is home to 10 notable supercomputers, the largest of which is the Pégaso¹⁰ cluster owned by Brazil oil giant Petrobras¹¹ with 19 petaflops, or 1.9e + 16 flops. This company owns six out of the ten supercomputers on the continent and mostly uses them for tasks such as geophysical exploration. Brazil is rapidly developing AI infrastructure, but the rest of South America lags behind partly due to political instability.

A major issue with the power of AI is that practitioners may become overly reliant on it, leading to automation bias. This can result in harmful outcomes like incorrect decisions, overdiagnosis, overtreatment, and defensive medicine (118). These issues arise because AI systems, despite being advanced, can still make mistakes or miss important nuances in a patient's condition. And so, the clinician's supervision is crucial for ensuring the best possible care. On the other hand, patients may also distrust AI and most of them are expressing discomfort with physicians who are solely relying on AI for medical care. This is mainly influenced by factors such as education, knowledge, and experiences on AI. For example, in pain management or surgical procedures, some patients are skeptical about using AI because they believe that it will only worsen their condition or offer no significant benefits. They are also reluctant in using AI chatbots for mental health support without a therapist's involvement because they believe that AI does not offer any emotional support. However, those people with higher education and more experiences in using AI generally view AI as a helpful tool for improving patient care. Thus, skepticisms may have arisen from limited understanding and data privacy concerns (119).

In addition, patients are more than just their health—they have unique needs, wishes, and values. Healthcare is built on a special relationship between doctors and patients, focused on respect, privacy, fairness, and care. This relationship should support patients' choices and protect them from harm. Thus, medical AI should align with these principles (118). Many medical schools present a fictional story about a woman who visits her doctor to receive her biopsy results. Turns out, she has metastatic cancer and was told she has only a few months to live. While the diagnosis and prognosis are precise, backed by copious evidence, it was revealed that the “doctor” delivering the news is actually an AI program with little emotional engagement. Although most oncologists do not fear being replaced by AI, some may be concerned that companies and healthcare professionals might fully rely on AI in the name of efficiency or profit. Thus, data scientists should be aware of the role they might play in this. As data scientists may become more involved in tumor boards, they should emphasize how AI can enhance the patient experience by ensuring they feel genuinely cared for.

Federated learning (FL) is a shared global model training framework that keeps data localized and addresses privacy concerns associated with sensitive and fragmented datasets (84, 121–124). Previously, model training has involved aggregating data from various centers all together, but in FL, each center completes part of the overall operation without its own data being accessible to other centers. Examples of implementation include improving hospital mortality prediction (125) and monitoring air pollution (126). As discussed, many medical and medical research centers in developing countries lack the infrastructure for high-bandwidth data transfers and processes with high computational cost. We see FL as a potential solution to this problem because the local operations conducted require less-intensive data transfers and lower computational costs. Nvidia¹⁴, a primary leader in the production of AI-capable hardware, has taken a keen interest in FL, which bodes well for LMICs as they partner with Nvidia to increase their computing infrastructure.

Low-cost AI models play a crucial role in addressing the unique challenges faced by LMICs, where resource constraints often limit access to advanced technologies. These models, designed to operate efficiently on low-power devices or with limited computational resources, have demonstrated significant potential across various domains. For instance, Zehra et al. (127) highlight strategies for implementing cost-effective AI-driven digital pathology, enabling earlier and more accessible diagnoses in resource-constrained healthcare systems. Similarly, Gangavarapu (128) introduces a multilingual medical AI model tailored to reduce health disparities in low-resource regions, emphasizing adaptability and equitable access. Beyond healthcare, initiatives like MAIScope (129), a portable AI-powered microscope, showcase how low-cost solutions can automate disease diagnoses in remote settings, bridging critical gaps in healthcare delivery. Divyashree et al. (130) aim to build an affordable healthcare system by integrating low-cost sensors, IoT, and AI to deliver real-time monitoring and diagnostic solutions, making healthcare more accessible and reliable for underserved populations. Such efforts underscore the necessity of designing AI models that not only meet technical requirements but also align with the socioeconomic realities of LMICs. As López et al. (131) argue, integrating these solutions into LMIC ecosystems requires collaboration and innovation to overcome barriers like data quality and infrastructure limitations, paving the way for sustainable AI adoption in underserved regions.

These examples further emphasize how low-cost AI models can address systemic challenges in LMICs. For instance, frugal machine learning (FML) emphasizes hardware, model, and data efficiency, enabling AI to be more accessible and sustainable without compromising performance (132). Similarly, field-programmable gate array (FPGA)-based reconfigurable clusters, as demonstrated by Rupanetti et al. (133), offer scalable and cost-effective solutions for machine learning applications in resource-constrained environments, improving parallel processing capabilities. Wearable devices have shown promise in reducing reliance on expensive ICU equipment, while leveraging AI for adaptive decision-making (130, 134). Such approaches not only enhance accessibility but also bridge gaps in healthcare delivery, ensuring that even under-resourced communities benefit from cutting-edge technology. These initiatives highlight the transformative potential of low-cost AI to improve quality of life, support local infrastructure, and promote equity in technology adoption worldwide.

To ensure the successful integration of data science in cancer care, coordinated efforts from multiple sectors are required:

Government Support: Governments must play an active role in prioritizing investments in health data systems and infrastructure. Policymakers can create enabling environments by establishing regulations that encourage innovation, ensuring data privacy and security, and providing funding for digital health initiatives (135).

Despite the challenges, the Global South has a unique opportunity to leapfrog traditional models of cancer care by adopting AI and data science technologies:

Leapfrogging Traditional Infrastructure: LMICs can bypass the need for costly physical infrastructure by adopting cloud-based solutions and mobile health (mHealth) applications. These technologies allow healthcare providers to access and analyze data remotely, making cancer care more scalable and accessible in underserved regions (34).

Adopting data science and artificial intelligence (AI) in healthcare, especially in LMICs, presents exciting opportunities for growth and development. Embracing these new methodologies can help healthcare professionals enhance their work, while also recognizing that change can be challenging. Several theories can guide this transition, supporting healthcare professionals as they explore new approaches and thoughtfully adjust established practices.

Transformational Learning Theory: Transformational learning, as outlined by Mezirow (138), offers a way to approach change with openness and reflection. In healthcare, this means taking the time to evaluate long-standing methods and consider how data-driven technologies might complement or enhance them. This theory emphasizes that by reflecting on their practices and outcomes, professionals can make thoughtful shifts toward integrating AI and data science. Transformational learning can help healthcare teams embrace these technologies as useful tools in improving patient care and outcomes, especially when applied within tumor boards where personalized, data-driven decisions can have a profound impact.

Capacity-building initiatives must begin with a thorough understanding of the specific challenges faced by LMIC healthcare systems. Surveys and focus groups should be conducted with tumor boards, healthcare providers, academic institutions, and government stakeholders to identify current gaps in skills, resources, and infrastructure. For example, many LMICs lack access to annotated clinical datasets, computational tools, and trained personnel for data-intensive tasks.

Building a consortium of stakeholders, including international academic institutions, regional healthcare organizations, and private sector partners, ensures the initiative is contextually relevant. Engagement with patient advocacy groups further aligns training programs with community-centered healthcare goals.

A foundational framework provides the competency benchmarks necessary for effective training and professional development. Drawing on global standards like the EDISON Data Science Competence Framework (43–46), this roadmap emphasizes interdisciplinary knowledge across oncology, data analytics, and machine learning. Specific competencies include tumor biology modeling, predictive analytics, and ethical considerations in data sharing.

Frameworks must also consider the sociocultural and economic contexts of LMICs, ensuring alignment with local healthcare priorities and available resources. Regional collaboration, such as through ASEAN or African Union initiatives, can drive the development of frameworks tailored to shared challenges.

Curriculum development is pivotal in equipping data scientists with both theoretical and practical skills. Academic partnerships can facilitate the co-creation of open-access curricula, emphasizing modular and flexible learning paths. Virtual training platforms like Coursera or edX, tailored for low-bandwidth environments, can provide asynchronous learning opportunities.

Hands-on training should be prioritized by integrating real-world case studies relevant to tumor board decision-making. Open-source computational tools and healthcare datasets, such as simulated tumor registries, enable practical skill acquisition even in resource-constrained settings.

Pilot programs serve as a testing ground for scaling training initiatives. These programs should be launched in select LMIC regions with strong healthcare and academic infrastructure, such as regional centers of excellence. Initial efforts could focus on equipping hospitals with basic computational resources and infrastructure and engaging local experts as trainers and mentors.

Collaborative public-private partnerships can address infrastructure gaps by providing access to cloud computing and storage solutions. Pilot sites must incorporate robust evaluation metrics to refine program delivery and outcomes.

Formal certification ensures that trained data scientists are recognized for their competencies, facilitating their integration into tumor boards. Developing a unified certification framework for specific regions [e.g., ASEAN, South Asian Association for Regional Cooperation (SAARC), African Union] enhances portability and consistency of credentials across borders. The creation of tiered certifications (basic, intermediate, advanced) allows for progressive skill development.

International accreditation bodies, such as ACM or WHO, can partner with regional institutions to standardize training programs. Accredited training centers within LMICs can sustain capacity-building efforts and encourage local leadership in data science education.

Scaling capacity-building efforts requires an iterative approach that incorporates feedback from pilot programs. Successful models can be expanded to additional LMICs, leveraging lessons learned to optimize training methodologies. Continuous engagement with stakeholders ensures curricula remain aligned with evolving healthcare and technological landscapes.

Regular updates to learning materials, such as new case studies or AI tools, maintain the relevance and quality of training programs. Regional and global funding initiatives must be actively pursued to sustain long-term capacity building.

Through this capacity-building roadmap initiative, LMICs can transform their healthcare systems into hubs of data-driven care, realizing the promise of equitable cancer treatment and research worldwide.