Developing Quantitative Models for Linking Radiation Treatment Options with Clinical Outcomes

Developing Quantitative Models for Linking Radiation Treatment Options with Clinical Outcomes Developing Quantitative Models for Linking Radiation Treatment Options with Clinical Outcomes Purpose: we are aiming at the development of capabilities to perform radiobiological modeling of treatment outcomes in radiation therapy, both in-house capability to clinical complications and local tumor control probability, using dosimetric and clinical outcomes data collected at Mayo Clinic. General Benefit: radiobiological modeling of a likelihood of treatment outcome is a tool that enables quantitative analysis of tradeoffs inherent in the practice of radiation therapy. Common examples of such tradeoffs are increasing treatment effectiveness (local control) at a cost of increasing likelihood of complications, or an expensive and invasive clinical intervention which can potentially reduce likelihood of clinical complications. Analysis of tradeoffs is an important part of the treatment planning process today, but such analysis becomes critical to providers and to insurance carriers when novel but costly treatment options become available. General Methodology: radiobiological modeling of clinical outcomes requires that a mathematical formula (a model) is developed which describes a relationship between dose distribution in an organ and a probability of a clinical outcome, such as local tumor control or a clinical complication. Because of our incomplete knowledge of biological mechanisms that lead to clinical outcomes, presently available models are parameterized which means that the mathematical formula includes a set of adjustable parameters with values that are not known a priori. Parameters of a model can only be determined from retrospective clinical outcomes data, through a procedure called model fitting. Model fitting is a numerical procedure that utilizes a database of patients for whom both the dosimetry (dose distributions in organs) and corresponding clinical outcomes were recorded. The procedure establishes a numerical measure of likelihood that a particular set of parameter values is consistent with observed clinical outcomes. The parameter space is systematically searched by a computer program until a parameter set is found that maximizes the likelihood function. Once parameters are constrained, the model can be used to predict the probability of an outcome for a given treatment plan prospectively. As more data is collected over time, the model can be progressively refined. It should be noted that radiobiological modeling has been practiced by the radiotherapy community for several decades. The most notable of these efforts is a research group called QUANTEC, which is a joint effort by ASTRO and AAPM. QUANTEC group published parameter sets for select models based on meta-analysis of existing literature. However, the work of QUANTEC was based on treatment techniques that are now outdated. Given uncertainties which are inherent in the models, it is critical that conclusions of the QUANTEC study are verified (either confirmed or updated) using clinical data obtained with newer treatment techniques. This need is particularly urgent for proton therapy, as dose distributions in proton therapy can be radically different from those seen in photon therapies. The QUANTEC group acknowledged this necessity in its publications, and explicitly called for continuing studies with newer treatment modalities. Specific Methodology: Model fitting using patient data at Mayo Clinic requires that new tools and processes being developed, most of which are specific to the institution. Dosimetry data has to be extracted from the database of Eclipse treatment planning system (Varian, Inc.) by an automated procedure and converted into a form which is suitable for further numerical processing. Clinical outcomes data has to be extracted from the hospital EMR, and from the departmental EMR. At present, extraction of clinical outcomes data is accomplished through reading of records by a qualified human researcher, which is a time consuming and expensive method. A set of tools is currently being developed by Mayo Clinic Rochester to automate this process to an extent possible. Lastly, tools have to be developed that combine dosimetric data and clinical outcomes data into a model fitting package according to the general methodology which was described in the previous paragraph.