Search NASA⌕ Search

SEARCH · Search NASA

Results for “Domain knowledge”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 127 records · Page 7

Leveraging data mining, active learning, and domain adaptation for efficient discovery of advanced oxygen evolution electrocatalysts

Developing advanced catalysts for acidic oxygen evolution reaction (OER) is crucial for sustainable hydrogen production. This study presents a multistage machine learning (ML) approach to streamline the discovery and optimization of complex multimetallic catalysts. Our method integrates data mining, active learning, and domain adaptation throughout the materials discovery process. Unlike traditional trial-and-error methods, this approach systematically narrows the exploration space using domain knowledge with minimized reliance on subjective intuition. Then, the active learning module efficiently refines element composition and synthesis conditions through iterative experimental feedback. The process culminated in the discovery of a promising Ru-Mn-Ca-Pr oxide catalyst. Our workflow also enhances theoretical simulations with domain adaptation strategy, providing deeper mechanistic insights aligned with experimental findings. By leveraging diverse data sources and multiple ML strategies, we demonstrate an efficient pathway for electrocatalyst discovery and optimization. This comprehensive, data-driven approach represents a paradigm shift and potentially benchmark in electrocatalysts research.

Science & Technology - Other Topics↗

Machine learning in materials science: From explainable predictions to autonomous design

The advent of big data and algorithmic developments in the field of machine learning (and artificial intelligence, in general) have greatly impacted the entire spectrum of physical sciences, including materials science. Materials data, measured or computed, combined with various techniques of machine learning have been employed to address a myriad of challenging problems, such as, development of efficient and predictive surrogate models for a range of materials properties, screening and down-selection of novel candidate materials for targeted applications, new methodologies to improve and further expedite molecular and atomistic simulations, with likely many more important developments to come in the foreseeable future. While the applications thus far have provided a glimpse of the true potential data-enabled routes have to offer, it has also become clear that further progress in this direction hinges on our ability to understand, explain and rationalize findings of a machine learning model in light of the domain-knowledge. This focused review provides an overview of the main areas where machine learning has been widely and successfully used in materials science. Subsequently, a brief discussion of several techniques that have been helpful in extracting physically-meaningful insights, causal relationships and design-centric knowledge from materials data is provided. Finally, we identify some of the imminent opportunities and challenges that materials community faces in this exciting and rapidly growing field.

36 MATERIALS SCIENCE↗

Advanced characterization-informed machine learning framework and quantitative insight to irradiated annular U-10Zr metallic fuels

Abstract U-10Zr Metal fuel is a promising nuclear fuel candidate for next-generation sodium-cooled fast spectrum reactors. Since the Experimental Breeder Reactor-II in the late 1960s, researchers accumulated a considerable amount of experience and knowledge on fuel performance at the engineering scale. However, a mechanistic understanding of fuel microstructure evolution and property degradation during in-reactor irradiation is still missing due to a lack of appropriate tools for rapid fuel microstructure assessment and property prediction based on post irradiation examination. This paper proposed a machine learning enabled workflow, coupled with domain knowledge and large dataset collected from advanced post-irradiation examination microscopies, to provide rapid and quantified assessments of the microstructure in two reactor irradiated prototypical annular metal fuels. Specifically, this paper revealed the distribution of Zr-bearing secondary phases and constitutional redistribution across different radial locations. Additionally, the ratios of seven different microstructures at various locations along the temperature gradient were quantified. Moreover, the distributions of fission gas pores on two types of U-10Zr annular fuels were quantitatively compared.

36 MATERIALS SCIENCE↗

A study of mapping exogenous knowledge representations into CONFIG

Qualitative reasoning is reasoning with a small set of qualitative values that is an abstraction of a larger and perhaps infinite set of quantitative values. The use of qualitative and quantitative reasoning together holds great promise for performance improvement in applications that suffer from large and/or imprecise knowledge domains. Included among these applications are the modeling, simulation, analysis, and fault diagnosis of physical systems. Several research groups continue to discover and experiment with new qualitative representations and reasoning techniques. However, due to the diversity of these techniques, it is difficult for the programs produced to exchange system models easily. The availability of mappings to transform knowledge from the form used by one of these programs to that used by another would open the doors for comparative analysis of these programs in areas such as completeness, correctness, and performance. A group at the Johnson Space Center (JSC) is working to develop CONFIG, a prototype qualitative modeling, simulation, and analysis tool for fault diagnosis applications in the U.S. space program. The availability of knowledge mappings from the programs produced by other research groups to CONFIG may provide savings in CONFIG's development costs and time, and may improve CONFIG's performance. The study of such mappings is the purpose of the research described in this paper. Two other research groups that have worked with the JSC group in the past are the Northwest University Group and the University of Texas at Austin Group. The former has produced a qualitative reasoning tool named SIMGEN, and the latter has produced one named QSIM. Another program produced by the Austin group is CC, a preprocessor that permits users to develop input for eventual use by QSIM, but in a more natural format. CONFIG and CC are both based on a component-connection ontology, so a mapping from CC's knowledge representation to CONFIG's knowledge representation was chosen as the focus of this study. A mapping from CC to CONFIG was developed. Due to differences between the two programs, however, the mapping transforms some of the CC knowledge to CONFIG as documentation rather than as knowledge in a form useful to computation. The study suggests that it may be worthwhile to pursue the mappings further. By implementing the mapping as a program, actual comparisons of computational efficiency and quality of results can be made between the QSIM and CONFIG programs. A secondary study may reveal that the results of the two programs augment one another, contradict one another, or differ only slightly. If the latter, the qualitative reasoning techniques may be compared in other areas, such as computational efficiency.

Mayfield, Blayne E.↗

Evaluation of Visualization Heuristics

Heuristics have been used since the early 1990's in the Human Computer Interaction (HCI) domain, for evaluating user interfaces. Most notably, heuristics developed by Nielsen and Molich (1990) have been used by the HCI community. Forsell and Johansson (2010) used the same methodology as Nielsen and Molich and developed a set of 10 heuristics for information visualization. Currently, neither the information visualization or the visual analytics community has adopted these heuristics. However, several studies have been conducted using heuristics. Hearst et al. (2015) found that using the heuristics for visualizations along with asking questions about the data resulted in complementary results as the questions about the data could compensate for heuristics that were difficult to understand and apply. Heli Väätäjä et al. (2016) found that heuristics related to interaction, veracity, and aesthetics needed to be added to the Forsell and Johansson (2010) set. They also noted that a lack of domain knowledge caused the evaluators issues in conducting a heuristic evaluation. We conducted a controlled experiment to determine what factors might impact the expert evaluators' capability to conduct a heuristic evaluation of a static information visualization. In particular, we wanted to understand how their familiarity with the visualizations and their experience in conducting heuristic evaluations impacted their evaluations of visualizations. We recruited 10 participants who were domain experts in the visualization field each having experience with one of more of the following: user experience design or research, user interface development, and human computer interaction research. We used a heuristic set proposed by Forsell and Johansson. Participants were also asked to complete a demographic survey in which they rated their experience with each visualization type generally, their experience conducting heuristic evaluations, their experience using each visualization type to extract information, and their experience developing and creating each visualization type. The participants were shown a series of 5 common visualizations (Scatterplots, Sunbursts, Tree Maps, Parallel Sets, and Area Graphs) all displaying the VAST 2008 Mini Challenge Two- Migrant Boats data . Participants were given a suggested analytical question related to each visualization to keep in mind as they applied the 10 heuristics to each visualization. They were asked if they could find any usability problems in the visualizations and which heuristic(s) helped them to identify these problems. The participants were asked to rate how relevant each heuristic was for evaluating each visualization along with rating how confident they felt in their ability to apply the heuristic to evaluate each visualization. At the conclusion of the study, participants provided qualitative feedback on the usability issues of each visualization. Additionally, we collected eye tracking data on 6 of the 10 participants. We will be using this data to determine how long participants looked at the various visualizations and what parts of the visualizations they focused on. We had two hypotheses for this empirical evaluation: 1.The Forsell and Johansson heuristics of Orientation and help, Data set reduction, and Prompting are less useful for static visualizations because they require some interaction in order to test completely. 2.Participants that do not have experience with heuristic evaluation will provide lower evaluation scores for the heuristics compared to those that have experience. Related to hypothesis 1, as our study used only static visualizations, interaction-related usability challenges should be harder to identify. This should result in lower usefulness ratings for those heuristics. Preliminary results indicate that participants found it more difficult to apply heuristics to unfamiliar visualizations.

Williams, Ryan G.↗

Using Induction to Refine Information Retrieval Strategies

Conceptual information retrieval systems use structured document indices, domain knowledge and a set of heuristic retrieval strategies to match user queries with a set of indices describing the document's content. Such retrieval strategies increase the set of relevant documents retrieved (increase recall), but at the expense of returning additional irrelevant documents (decrease precision). Usually in conceptual information retrieval systems this tradeoff is managed by hand and with difficulty. This paper discusses ways of managing this tradeoff by the application of standard induction algorithms to refine the retrieval strategies in an engineering design domain. We gathered examples of query/retrieval pairs during the system's operation using feedback from a user on the retrieved information. We then fed these examples to the induction algorithm and generated decision trees that refine the existing set of retrieval strategies. We found that (1) induction improved the precision on a set of queries generated by another user, without a significant loss in recall, and (2) in an interactive mode, the decision trees pointed out flaws in the retrieval and indexing knowledge and suggested ways to refine the retrieval strategies.

Baudin, Catherine↗

Automatic Debugging Support for UML Designs

Design of large software systems requires rigorous application of software engineering methods covering all phases of the software process. Debugging during the early design phases is extremely important, because late bug-fixes are expensive. In this paper, we describe an approach which facilitates debugging of UML requirements and designs. The Unified Modeling Language (UML) is a set of notations for object-orient design of a software system. We have developed an algorithm which translates requirement specifications in the form of annotated sequence diagrams into structured statecharts. This algorithm detects conflicts between sequence diagrams and inconsistencies in the domain knowledge. After synthesizing statecharts from sequence diagrams, these statecharts usually are subject to manual modification and refinement. By using the "backward" direction of our synthesis algorithm. we are able to map modifications made to the statechart back into the requirements (sequence diagrams) and check for conflicts there. Fed back to the user conflicts detected by our algorithm are the basis for deductive-based debugging of requirements and domain theory in very early development stages. Our approach allows to generate explanations oil why there is a conflict and which parts of the specifications are affected.

Schumann, Johann↗

SciDAC's Earth System Grid Center for Enabling Technologies Semiannual Progress Report October 1, 2010 through March 31, 2011

This report summarizes work carried out by the Earth System Grid Center for Enabling Technologies (ESG-CET) from October 1, 2010 through March 31, 2011. It discusses ESG-CET highlights for the reporting period, overall progress, period goals, and collaborations, and lists papers and presentations. To learn more about our project and to find previous reports, please visit the ESG-CET Web sites: http://esg-pcmdi.llnl.gov/ and/or https://wiki.ucar.edu/display/esgcet/Home. This report will be forwarded to managers in the Department of Energy (DOE) Scientific Discovery through Advanced Computing (SciDAC) program and the Office of Biological and Environmental Research (OBER), as well as national and international collaborators and stakeholders (e.g., those involved in the Coupled Model Intercomparison Project, phase 5 (CMIP5) for the Intergovernmental Panel on Climate Change (IPCC) 5th Assessment Report (AR5); the Community Earth System Model (CESM); the Climate Science Computational End Station (CCES); SciDAC II: A Scalable and Extensible Earth System Model for Climate Change Science; the North American Regional Climate Change Assessment Program (NARCCAP); the Atmospheric Radiation Measurement (ARM) program; the National Aeronautics and Space Administration (NASA), the National Oceanic and Atmospheric Administration (NOAA)), and also to researchers working on a variety of other climate model and observation evaluation activities. The ESG-CET executive committee consists of Dean N. Williams, Lawrence Livermore National Laboratory (LLNL); Ian Foster, Argonne National Laboratory (ANL); and Don Middleton, National Center for Atmospheric Research (NCAR). The ESG-CET team is a group of researchers and scientists with diverse domain knowledge, whose home institutions include eight laboratories and two universities: ANL, Los Alamos National Laboratory (LANL), Lawrence Berkeley National Laboratory (LBNL), LLNL, NASA/Jet Propulsion Laboratory (JPL), NCAR, Oak Ridge National Laboratory (ORNL), Pacific Marine Environmental Laboratory (PMEL)/NOAA, Rensselaer Polytechnic Institute (RPI), and University of Southern California, Information Sciences Institute (USC/ISI). All ESG-CET work is accomplished under DOE open-source guidelines and in close collaboration with the project's stakeholders, domain researchers, and scientists. Through the ESG project, the ESG-CET team has developed and delivered a production environment for climate data from multiple climate model sources (e.g., CMIP (IPCC), CESM, ocean model data (e.g., Parallel Ocean Program), observation data (e.g., Atmospheric Infrared Sounder, Microwave Limb Sounder), and analysis and visualization tools) that serves a worldwide climate research community. Data holdings are distributed across multiple sites including LANL, LBNL, LLNL, NCAR, and ORNL as well as unfunded partners sites such as the Australian National University (ANU) National Computational Infrastructure (NCI), the British Atmospheric Data Center (BADC), the Geophysical Fluid Dynamics Laboratory/NOAA, the Max Planck Institute for Meteorology (MPI-M), the German Climate Computing Centre (DKRZ), and NASA/JPL. As we transition from development activities to production and operations, the ESG-CET team is tasked with making data available to all users who want to understand it, process it, extract value from it, visualize it, and/or communicate it to others. This ongoing effort is extremely large and complex, but it will be incredibly valuable for building 'science gateways' to critical climate resources (such as CESM, CMIP5, ARM, NARCCAP, Atmospheric Infrared Sounder (AIRS), etc.) for processing the next IPCC assessment report. Continued ESG progress will result in a production-scale system that will empower scientists to attempt new and exciting data exchanges, which could ultimately lead to breakthrough climate science discoveries.

54 ENVIRONMENTAL SCIENCES↗

Data Fusion via Neural Network Entropy Minimization for Target Detection and Multi-Sensor Event Classification

Broadly applicable solutions to multimodal and multisensory fusion problems across domains remain a challenge because effective solutions often require substantive domain knowledge and engineering. The chief questions that arise for data fusion are in when to share information from different data sources, and how to accomplish the integration of information. The solutions explored in this work remain agnostic to input representation and terminal decision fusion approaches by sharing information through the learning objective as a compound objective function. The objective function this work uses assumes a one-to-one learning paradigm within a one-to-many domain which allows the assumption that consistency can be enforced across the one-to-many dimension. The domains and tasks we explore in this work include multi-sensor fusion for seismic event location and multimodal hyperspectral target discrimination. We find that our domain- informed consistency objectives are challenging to implement in stable and successful learning because of intersections between inherent data complexity and practical parameter optimization. While multimodal hyperspectral target discrimination was not enhanced across a range of different experiments by the fusion strategies put forward in this work, seismic event location benefited substantially, but only for label-limited scenarios.

70 PLASMA PHYSICS AND FUSION TECHNOLOGY↗

Goal Directed Model Inversion: Learning Within Domain Constraints

Goal Directed Model Inversion (GDMI) is an algorithm designed to generalize supervised learning to the case where target outputs are not available to the learning system. The output of the learning system becomes the input to some external device or transformation, and only the output of this device or transformation can be compared to a desired target. The fundamental driving mechanism of GDMI is to learn from success. Given that a wrong outcome is achieved, one notes that the action that produced that outcome "would have been right if the outcome had been the desired one." The algorithm makes use of these intermediate "successes" to achieve the final goal. A unique and potentially very important feature of this algorithm is the ability to modify the output of the learning module to force upon it a desired syntactic structure. This differs from ordinary supervised learning in the following way: in supervised learning the exact desired output pattern must be provided. In GDMI instead, it is possible to require simply that the output obey certain rules, i.e., that it "make sense" in some way determined by the knowledge domain. The exact pattern that will achieve the desired outcome is then found by the system. The ability to impose rules while allowing the system to search for its own answers in the context of neural networks is potentially a major breakthrough in two ways: 1) it may allow the construction of networks that can incorporate immediately some important knowledge, i.e. would not need to learn everything from scratch as normally required at present, and 2) learning and searching would be limited to the areas where it is necessary, thus facilitating and speeding up the process. These points are illustrated with examples from robotic path planning and parametric design.

Colombano, Silvano P.↗

Goal Directed Model Inversion: Learning Within Domain Constraints

Goal Directed Model Inversion (GDMI) is an algorithm designed to generalize supervised learning to the case where target outputs are not available to the learning system. The output of the learning system becomes the input to some external device or transformation, and only the output of this device or transformation can be compared to a desired target. The fundamental driving mechanism of GDMI is to learn from success. Given that a wrong outcome is achieved, one notes that the action that produced that outcome "would have been right if the outcome had been the desired one." The algorithm makes use of these intermediate "successes" to achieve the final goal. A unique and potentially very important feature of this algorithm is the ability to modify the output of the learning module to force upon it a desired syntactic structure. This differs from ordinary supervised learning in the following way: in supervised learning the exact desired output pattern must be provided. In GDMI instead, it is possible to require simply that the output obey certain rules, i.e., that it "make sense" in some way determined by the knowledge domain. The exact pattern that will achieve the desired outcome is then found by the system. The ability to impose rules while allowing the system to search for its own answers in the context of neural networks is potentially a major breakthrough in two ways: (1) it may allow the construction of networks that can incorporate immediately some important knowledge, i.e., would not need to learn everything from scratch as normally required at present; and (2) learning and searching would be limited to the areas where it is necessary, thus facilitating and speeding up the process. These points are illustrated with examples from robotic path planning and parametric design.

Colombano, Silvano P.↗

Machine learning with knowledge constraints for process optimization of open-air perovskite solar cell manufacturing

Perovskite photovoltaics (PV) have achieved rapid development in the past decade in terms of power conversion efficiency of small-area lab-scale devices; however, successful commercialization still requires further development of low-cost, scalable, and high-throughput manufacturing techniques. One of the critical challenges of developing a new fabrication technique is the high-dimensional parameter space for optimization, but machine learning (ML) can readily be used to accelerate perovskite PV scaling. Herein, we present an ML-guided framework of sequential learning for manufacturing process optimization. We apply our methodology to the Rapid Spray Plasma Processing (RSPP) technique for perovskite thin films in ambient conditions. With a limited experimental budget of screening 100 process conditions, we demonstrated an efficiency improvement to 18.5% as the best-in-our-lab device fabricated by RSPP, and we also experimentally found 10 unique process conditions to produce the top-performing devices of more than 17% efficiency, which is 5 times higher rate of success than the control experiments with pseudo-random Latin hypercube sampling. Our model is enabled by three innovations: (a) flexible knowledge transfer between experimental processes by incorporating data from prior experimental data as a probabilistic constraint; (b) incorporation of both subjective human observations and ML insights when selecting next experiments; (c) adaptive strategy of locating the region of interest using Bayesian optimization first, and then conducting local exploration for high-efficiency devices. Furthermore, in virtual benchmarking, our framework achieves faster improvements with limited experimental budgets than traditional design-of-experiments methods (e.g., one-variable-at-a-time sampling). This framework shows the capability of incorporating researchers’ domain knowledge into the ML-guided optimization loop; therefore, it has the potential to facilitate the wider adoption of ML in scaling to perovskite PV manufacturing.

14 SOLAR ENERGY↗

Using Rule-Based Models for Weak Supervised Learning: A ChemNet for Transferable Chemical Property Prediction

With access to large datasets, deep neural networks (DNN) have achieved human-level accuracy in image and speech recognition tasks. However, in chemistry, data is inherently small and fragmented. In this work, we develop an approach of integrating rule-based knowledge into the Chemception CNN model via transfer learning techniques. The resulting model, ChemNet, is a transferable and generalizable deep neural network for chemical property prediction that learns in a semi-supervised manner from large unlabeled chemical databases. When ChemNet is further fine-tuned on 3 smaller datasets to predict chemical properties that it was not originally trained on, we show that ChemNet exceeds the accuracy of existing Chemception CNN models and other contemporary DNN models that were trained using conventional supervised learning approaches. These results indicate that pre-training ChemNet on a large diverse chemical database while incorporating chemistry domain knowledge, enables the development of more generalizable deep neural networks for the prediction of novel chemical properties.

Artificial Intelligence, Cheminformatics, Computat↗

A Tale from the Trenches: Applying Metamorphic and Differential Testing to Bioinformatics Software

Metamorphic and differential testing have been proposed as best practices for testing software that is difficult to test, such as for programs in scientific domains. An assumption is that these approaches can be easily customized and applied to almost any domain. However, scientific software is often data-driven, and metamorphic relations may require significant domain knowledge to develop. In addition, tools are often written for ad-hoc experimentation by the scientists and often embed many assumptions about the importance and representation of different natural phenomena. In this paper, we present our experience applying both metamorphic and differential testing to a set of four computational biology tools that predict the growth of an organism. While our original goal was to evaluate these techniques to improve our system-level testing, we encountered multiple roadblocks along the way. Although we did find faults (some confirmed by developers), we also uncovered a set of challenges, including the considerable manual effort required for (a) defining domain-specific tests, (b) validating correctness, and (c) distinguishing between issues stemming from poor data and those arising from incorrect software.

Marsh, Alexis L [Iowa State University/Ames Labora↗

What Machine Learning Can and Cannot Do for Inertial Confinement Fusion

Machine learning methodologies have played remarkable roles in solving complex systems with large data, well-defined input–output pairs, and clearly definable goals and metrics. The methodologies are effective in image analysis, classification, and systems without long chains of logic. Recently, machine-learning methodologies have been widely applied to inertial confinement fusion (ICF) capsules and the design optimization of OMEGA (Omega Laser Facility) capsule implosion and NIF (National Ignition Facility) ignition capsules, leading to significant progress. As machine learning is being increasingly applied, concerns arise regarding its capabilities and limitations in the context of ICF. ICF is a complicated physical system that relies on physics knowledge and human judgment to guide machine learning. Additionally, the experimental database for ICF ignition is not large enough to provide credible training data. Most researchers in the field of ICF use simulations, or a mix of simulations and experimental results, instead of real data to train machine learning models and related tools. They then use the trained learning model to predict future events. This methodology can be successful, subject to a careful choice of data and simulations. However, because of the extreme sensitivity of the neutron yield to the input implosion parameters, physics-guided machine learning for ICF is extremely important and necessary, especially when the database is small, the uncertain-domain knowledge is large, and the physical capabilities of the learning models are still being developed. In this work, we identify problems in ICF that are suitable for machine learning and circumstances where machine learning is less likely to be successful. This study investigates the applications of machine learning and highlights fundamental research challenges and directions associated with machine learning in ICF.

70 PLASMA PHYSICS AND FUSION TECHNOLOGY↗

SNAD transient miner: Finding missed transient events in ZTF DR4 using k-D trees

Here we report the automatic detection of 11 transients (7 possible supernovae and 4 active galactic nuclei candidates) within the Zwicky Transient Facility fourth data release (ZTF DR4), all of them observed in 2018 and absent from public catalogs. Among these, three were not part of the ZTF alert stream. Our transient mining strategy employs 41 physically motivated features extracted from both real light curves and four simulated light curve models (SN Ia, SN II, TDE, SLSN-I). These features are input to a k-D tree algorithm, from which we calculate the 15 nearest neighbors. After pre-processing and selection cuts, our dataset contained approximately a million objects among which we visually inspected the 105 closest neighbors from seven of our brightest, most well-sampled simulations, comprising 89 unique ZTF DR4 sources. Our result illustrates the potential of coherently incorporating domain knowledge and automatic learning algorithms, which is one of the guiding principles directing the SNAD team. It also demonstrates that the ZTF DR is a suitable testing ground for data mining algorithms aiming to prepare for the next generation of astronomical data.

79 ASTRONOMY AND ASTROPHYSICS↗

Dimensions of complexity in learning from interactive instruction

Robot systems deployed in space must exhibit flexibility. In particular, an intelligent robotic agent should not have to be reprogrammed for each of the various tasks it may face during the course of its lifetime. However, pre-programming knowledge for all of the possible tasks that may be needed is extremely difficult. Therefore, a powerful notion is that of an instructible agent, one which is able to receive task-level instructions and advice from a human advisor. An agent must do more than simply memorize the instructions it is given (this would amount to programming). Rather, after mapping instructions into task constructs that it can reason with, it must determine each instruction's proper scope of applicability. In this paper, we will examine the characteristics of instruction, and the characteristics of agents, that affect learning from instruction. We find that in addition to a myriad of linguistic concerns, both the situatedness of the instructions (their placement within the ongoing execution of tasks) and the prior domain knowledge of the agent have an impact on what can be learned.

Huffman, Scott B.↗

Comparing Individualized Survival Predictions From Random Survival Forests and Multistate Models in the Presence of Missing Data: A Case Study of Patients With Oropharyngeal Cancer

Background: In recent years, interest in prognostic calculators for predicting patient health outcomes has grown with the popularity of personalized medicine. These calculators, which can inform treatment decisions, employ many different methods, each of which has advantages and disadvantages. Methods: We present a comparison of a multistate model (MSM) and a random survival forest (RSF) through a case study of prognostic predictions for patients with oropharyngeal squamous cell carcinoma. The MSM is highly structured and takes into account some aspects of the clinical context and knowledge about oropharyngeal cancer, while the RSF can be thought of as a black-box non-parametric approach. Key in this comparison are the high rate of missing values within these data and the different approaches used by the MSM and RSF to handle missingness. Results: We compare the accuracy (discrimination and calibration) of survival probabilities predicted by both approaches and use simulation studies to better understand how predictive accuracy is influenced by the approach to (1) handling missing data and (2) modeling structural/disease progression information present in the data. We conclude that both approaches have similar predictive accuracy, with a slight advantage going to the MSM. Conclusions: Although the MSM shows slightly better predictive ability than the RSF, consideration of other differences are key when selecting the best approach for addressing a specific research question. These key differences include the methods’ ability to incorporate domain knowledge, and their ability to handle missing data as well as their interpretability, and ease of implementation. Ultimately, selecting the statistical method that has the most potential to aid in clinical decisions requires thoughtful consideration of the specific goals.

60 APPLIED LIFE SCIENCES↗