Volume 32, Issue 4, October 2022, Pages 351-364

Outcome modeling plays an important role in personalizing radiotherapy and finds applications in specialized areas such as adaptive radiotherapy. Conventional outcome models that are based on a simplified understanding of radiobiological effects or empirical fitting often only consider dosimetric information. However, it is recognized that response to radiotherapy is multi-factorial and involves a complex interaction of radiation therapy, patient and treatment factors, and the tumor microenvironment. Recently, large pools of patient-specific biological and imaging data have become available with the development of advanced biotechnology and multi-modality imaging techniques. Given this complexity, artificial intelligence (AI) and machine learning (ML) are valuable to make sense of such a plethora of heterogeneous data and to aid clinicians in their decision-making process. The role of AI/ML has been demonstrated in many retrospective studies and more recently prospective evidence has been emerging as well to support AI/ML for personalized and precision radiotherapy.

Section snippetsWhat is Outcome Modeling?

In general, outcome modeling1 is a process of generating an abstract representation of input data and identifying the underlying relationship between these data and observed clinical outcomes. Outcome modeling has continuously evolved from trial-and-error approaches with back of the hand calculations into the application of advanced statistical and computing techniques2,3 such as machine learning (ML). ML is the study of computer algorithms that can learn from data without explicitly being

Relevance to Radiotherapy: Treatment Planning, Personalization, Adaptation

Outcome models19 could be used during the consultation period as a guide for discerning and determining treatment options by both the clinician and the patient. Alternatively, once a decision has been reached, planners can use TCP/NTCP models to drive the generation of personalized treatment plans20 during plan optimization, to maximize tumor eradication and minimize complication risk using competing risk metrics as the complication-free tumor control (p+) or others.21 Some example applications

Analytical Models

Conventional analytical models1 that are mainly based on dosimetric variables can be categorized as mechanistic or phenomenological. Mechanistic models mathematically formulate outcomes based on the simplified radiobiological understanding of radiation effects. The well-known linear quadratic (LQ) model33 and its derived biological effective dose (BED) are widely adopted to convert dose under different fractionation scheme. Phenomenological models fit the available dosimetric data to a

AI/ML Methods for Outcome Modeling

In most current systems, prediction of the outcome can be defined as a supervised learning problem in the context of ML. The goal of supervised learning is to infer a mapping function between the features x(input) and labels Y (output) based on a dataset. Specifically, in the outcome model, a mapping function is to be discovered between patient-specific information (input) and clinical endpoints (output):f(x,w):x→Y

After an outcome model is trained, the learned function can be then applied to an

Retrospective Validation

Retrospective validation62 evaluates the outcome model on the past treatments in patients who were not preselected for the treatment regimen in question, for example, those recommended by the outcome models. It can provide evidence to base a change in clinical practice and provide an insight into designing a new clinical protocol or prospective study.

The retrospective validation can be conducted internally, based on a single dataset, or externally, based on a separate study or

Steps Involved in AI-based Outcome Modeling

Similar to published checklists for example, MI-CLAIM66 for clinical research, CLAIM67 for radiology, and CLAMP62 for medical physics, steps and caveats involved in building a AI-based outcome model are summarized in Table 1. Given the complexity of the outcome model building process, the importance of its development and validation cannot be overemphasized. Multisteps of training and tuning should be appropriately adopted during the development of these models. Additionally, the

Breast cancer – Prediction of Acute Skin Toxicity

Thermal imaging of treated breast was adopted to predict acute skin toxicity in breast cancer patients who received adjuvant whole-breast RT (4250cGy/16 fractions).76 Quantitative imaging biomarkers including 26 surface temperature and texture features were extracted from thermal images of the treated breast before RT, after delivery of 5, 10, and 15 fractions.

A whole dataset of 90 patients was randomly split into a training (83%) and test dataset (17%), which is considered to be TRIPOD level-2

Outlook

AI and ML are expected to be the main driving approaches for developing outcome models in radiotherapy for the purpose of personalizing and/or adapting radiotherapy prescription among many other possible applications as summarized in Table 1. However, such an approach is prone to overfitting issues given the limited sample size data available, hence, rigorous model building, and validation plans are required as suggested in the checklist of Table 2. Institutional prospective evaluation studies

Conclusions

Outcome modeling generates a representation of patient data and correlates these data with observed clinical outcomes. It has a wide application in radiation oncology, including radiation treatment planning, adaptive and personalized treatments. In the era of big data, it is expected that AI-based outcome modeling empowered by patient-specific clinical, treatment, biological, dosimetric, and imaging data will pave the way for better understanding cancer biology and improving patients’ response.

View full text

© 2022 Elsevier Inc. All rights reserved.