Interpretable Machine Learning for Prediction with Missing Values at Test Time
Doctoral thesis, 2025
This thesis addresses the challenge of making interpretable predictions when feature values may be missing at deployment (at "test time''). Although imputation is a common strategy for handling missing values, it can obscure the relationship between inputs and predictions, thereby undermining interpretability and trust -especially in safety-critical domains such as healthcare. Alternatively, incorporating missingness indicators may introduce complexity and further reduce model interpretability. Tree-based models can handle missing values natively but are limited to specific model classes, potentially restricting flexibility and generalizability. To overcome these limitations, this thesis develops methods that (i) retain or improve predictive performance, (ii) handle missing values effectively at test time, and (iii) produce models that are simple and interpretable.
We first leverage missingness patterns by introducing Sharing Pattern Submodels, where a separate interpretable submodel is trained for each unique missingness pattern, with parameters shared across submodels via sparsity to enhance generalization. Next, we investigate training models that rarely require the values of missing (or imputed) features at test time. We introduce MINTY, a linear rule-based model that avoids imputation by allowing logical substitutions for missing features. We then generalize this idea through a missingness-avoiding framework, which extends to multiple model classes, including decision trees, sparse linear models, and ensembles, by incorporating classifier-specific regularization terms into their learning objectives to discourage reliance on missing values. To support the development of clinically valuable models, we conducted a clinician survey revealing that medical professionals favor models that natively handle missingness. Finally, we explore interpretable patient history representations for modeling policies in sequential clinical decision-making, shifting the focus from missingness to temporal modeling. Collectively, this work establishes methods for interpretable machine learning with test-time missingness, supported by both technical innovations and human-centered insights, to enable transparent and practical decision support.
We first leverage missingness patterns by introducing Sharing Pattern Submodels, where a separate interpretable submodel is trained for each unique missingness pattern, with parameters shared across submodels via sparsity to enhance generalization. Next, we investigate training models that rarely require the values of missing (or imputed) features at test time. We introduce MINTY, a linear rule-based model that avoids imputation by allowing logical substitutions for missing features. We then generalize this idea through a missingness-avoiding framework, which extends to multiple model classes, including decision trees, sparse linear models, and ensembles, by incorporating classifier-specific regularization terms into their learning objectives to discourage reliance on missing values. To support the development of clinically valuable models, we conducted a clinician survey revealing that medical professionals favor models that natively handle missingness. Finally, we explore interpretable patient history representations for modeling policies in sequential clinical decision-making, shifting the focus from missingness to temporal modeling. Collectively, this work establishes methods for interpretable machine learning with test-time missingness, supported by both technical innovations and human-centered insights, to enable transparent and practical decision support.
Missing Values
Decision Making
Healthcare
Machine learning
Interpretability
Room EE (6233), EDIT Building Hörsalsvägen 11, Chalmers University of Technology, Campus Johanneberg
Opponent: Prof. Stefan Feuerriegel, Institute of Artificial Intelligence (AI) in Management, LMU Munich School of Management, Germany
Author
Lena Stempfle
Data Science and AI 3
Prediction Models That Learn to Avoid Missing Values
Proceedings of Machine Learning Research,;Vol. 267(2025)
Paper in proceeding
Sharing Pattern Submodels for Prediction with Missing Values
Proceedings of the AAAI Conference on Artificial Intelligence,;Vol. 37(2023)p. 9882-9890
Paper in proceeding
MINTY: Rule-based Models that Minimize the Need for Imputing Features with Missing Values
Proceedings of the 38th Conference on Uncertainty in Artificial Intelligence, UAI 2022,;Vol. 238(2024)p. 964-972
Paper in proceeding
How Should We Represent History in Interpretable Models of Clinical Policies?
Proceedings of Machine Learning Research,;Vol. 259(2024)p. 714-734
Paper in proceeding
How can we make interpretable and accurate predictions when some information is missing?
In safety-critical fields like healthcare, clinicians often must make decisions based on incomplete records–lab results may be delayed, or certain tests never ordered. A common strategy to deal with missing data is imputation, where gaps are filled using averages, zeros, or statistical estimates. However, this can introduce bias and make predictions harder to interpret and trust.
This thesis explores new approaches to building models that work reliably even when some data is missing, without relying on imputation. The goal is to maintain or improve predictive performance while keeping the models interpretable and aligned with how clinicians reason. One method trains separate models for different patterns of missing data, for example, grouping patients with similar available measurements, while still sharing useful information across groups to generalize well. Another key contribution is a rule-based method that uses logical substitutions instead of imputing missing values. This idea is extended into a broader framework that helps models, like decision trees, sparse linear models, and ensembles, learn to avoid relying on imputed data altogether. This can help clinicians make better decisions using fewer additional measurements by leveraging frequently observed features, ultimately reducing delays and healthcare costs. These models, tested on diverse healthcare datasets, demonstrate strong predictive performance while reducing reliance on filled-in data. A clinician survey confirms a preference for models that explicitly recognize and handle missingness. The thesis also investigates interpretable models for time-series data, common in patient records, to support evolving treatment decisions over time.
Together, these contributions offer practical tools for building decision-making systems that are both interpretable and trustworthy–especially in clinical environments where missing data is the norm, not the exception.
In safety-critical fields like healthcare, clinicians often must make decisions based on incomplete records–lab results may be delayed, or certain tests never ordered. A common strategy to deal with missing data is imputation, where gaps are filled using averages, zeros, or statistical estimates. However, this can introduce bias and make predictions harder to interpret and trust.
This thesis explores new approaches to building models that work reliably even when some data is missing, without relying on imputation. The goal is to maintain or improve predictive performance while keeping the models interpretable and aligned with how clinicians reason. One method trains separate models for different patterns of missing data, for example, grouping patients with similar available measurements, while still sharing useful information across groups to generalize well. Another key contribution is a rule-based method that uses logical substitutions instead of imputing missing values. This idea is extended into a broader framework that helps models, like decision trees, sparse linear models, and ensembles, learn to avoid relying on imputed data altogether. This can help clinicians make better decisions using fewer additional measurements by leveraging frequently observed features, ultimately reducing delays and healthcare costs. These models, tested on diverse healthcare datasets, demonstrate strong predictive performance while reducing reliance on filled-in data. A clinician survey confirms a preference for models that explicitly recognize and handle missingness. The thesis also investigates interpretable models for time-series data, common in patient records, to support evolving treatment decisions over time.
Together, these contributions offer practical tools for building decision-making systems that are both interpretable and trustworthy–especially in clinical environments where missing data is the norm, not the exception.
Areas of Advance
Information and Communication Technology
Subject Categories (SSIF 2025)
Probability Theory and Statistics
Computer Sciences
Computer Systems
Roots
Basic sciences
Infrastructure
Chalmers e-Commons (incl. C3SE, 2020-)
ISBN
978-91-8103-246-8
Doktorsavhandlingar vid Chalmers tekniska högskola. Ny serie: 5704
Publisher
Chalmers
Room EE (6233), EDIT Building Hörsalsvägen 11, Chalmers University of Technology, Campus Johanneberg
Opponent: Prof. Stefan Feuerriegel, Institute of Artificial Intelligence (AI) in Management, LMU Munich School of Management, Germany