Source-linked AI summary
Handling missing values in healthcare data: A systematic review of deep learning-based imputation techniques
Mingxuan Liu, Siqi Li, Han Yuan, Marcus Eng Hock Ong, Yilin Ning, Feng Xie, Seyed Ehsan Saffari, Victor Volovici, Bibhas Chakraborty, Nan Liu
TL;DR
Missing values threaten reliable healthcare analysis as data become more diverse and complex. This systematic review mapped DL-based imputation across healthcare data types, model backbones, strategies, and comparator methods, finding that models can be customized to data types and missing patterns and that integrated imputation may enhance performance. The review was limited to clinical and translational research.
Problem
Diverse healthcare data create missing-value challenges, while existing reviews do not systematically and specifically evaluate DL-based imputation across data types.
Method
The authors systematically reviewed DL-based healthcare imputation studies and mapped data types, model backbones, imputation strategies, and comparisons with non-DL methods.
Results
DL models could be customized to data types and corresponding missing patterns, while integrated imputation could enhance performance.
Takeaways & Limitations
DL-based imputation offers a data-type-sensitive approach, with integrated strategies particularly relevant to complex healthcare data.
Takeaways & Limitations
The review was limited to clinical and translational research, so DL-based imputation techniques in other areas may not be represented.
Abstract
from arXiv · showhide
Objective: The proper handling of missing values is critical to delivering reliable estimates and decisions, especially in high-stakes fields such as clinical research. The increasing diversity and complexity of data have led many researchers to develop deep learning (DL)-based imputation techniques. We conducted a systematic review to evaluate the use of these techniques, with a particular focus on data types, aiming to assist healthcare researchers from various disciplines in dealing with missing values. Methods: We searched five databases (MEDLINE, Web of Science, Embase, CINAHL, and Scopus) for articles published prior to August 2021 that applied DL-based models to imputation. We assessed selected publications from four perspectives: health data types, model backbone (i.e., main architecture), imputation strategies, and comparison with non-DL-based methods. Based on data types, we created an evidence map to illustrate the adoption of DL models. Results: We included 64 articles, of which tabular static (26.6%, 17/64) and temporal data (37.5%, 24/64) were the most frequently investigated. We found that model backbone(s) differed among data types as well as the imputation strategy. The "integrated" strategy, that is, the imputation task being solved concurrently with downstream tasks, was popular for tabular temporal (50%, 12/24) and multi-modal data (71.4%, 5/7), but limited for other data types. Moreover, DL-based imputation methods yielded better imputation accuracy in most studies, compared with non-DL-based methods. Conclusion: DL-based imputation models can be customized based on data type, addressing the corresponding missing patterns, and its associated "integrated" strategy can enhance the efficacy of imputation, especially in scenarios where data is complex. Future research may focus on the portability and fairness of DL-based models for healthcare data imputation.
1 Centre for Quantitative Medicine, Duke-NUS Medical School, Singapore
The listed affiliation is the Department of Neurosurgery at Erasmus MC University Medical Center in Rotterdam, The Netherlands.
- The Department of Neurosurgery is affiliated with Erasmus MC University Medical Center in Rotterdam, The Netherlands.
8 Institute of Data Science, National University of Singapore, Singapore
The passage identifies equal author contribution and names Nan Liu as the corresponding author at Duke-NUS Medical School's Centre for Quantitative Medicine.
- These authors contributed equally.
- Nan Liu is the corresponding author at the Centre for Quantitative Medicine, Duke-NUS Medical School.
1. Introduction
Healthcare data are increasingly diverse, and newer formats are more susceptible to missing values than traditional tabular clinical data. The review motivates DL-based approaches because missingness can impair prediction and bias downstream healthcare decisions.
- Newer healthcare data formats, including temporal, imaging, and sequencing data, are more susceptible to missing values than traditional tabular clinical data.
- Missing values can reduce prediction power and bias downstream decision-making, posing particular problems in healthcare.
- Simple imputation methods may be unsuitable across diverse data types and architectures, while some methods assume missingness is MAR.
- DL methods have been used to improve imputation accuracy and can be tailored to complex missing patterns and structures such as sequential and spatial data.
- Existing reviews largely focus on non-DL methods or treat neural networks as one method category, limiting guidance for researchers applying DL models.
- This review addresses the gap by systematically mapping DL-based imputation methods across diverse healthcare data types and examining model use by data type.
2. Materials and Methods
The study used a PRISMA-guided systematic review of DL-based healthcare imputation research. Articles were categorized by data type, model backbone, imputation strategy, and comparisons with non-DL methods.
- The review searched MEDLINE, Web of Science, Embase, CINAHL, and Scopus for eligible healthcare imputation studies.
- Studies were screened independently by two reviewers, with disagreements resolved through consultation with a third reviewer.
- The analysis extracted data types, model backbones, imputation strategies, and comparisons with non-DL-based methods.
- Data types: Healthcare data were classified into tabular static, tabular temporal, genetic and genomic, image, signal, and multi-modal categories.
- Model backbones: Model backbones included MLPs, RNNs, autoencoders, GANs, hybrids, and less common models such as SOMs.
- Imputation strategies: Imputation strategies were classified as separated when distinct from downstream tasks and integrated when performed concurrently with them.
3. Results
Across 64 studies, DL-based imputation research covered multiple healthcare data types, with tabular temporal and static data most frequently investigated. Architectures and imputation strategies varied by data type, while separated strategies predominated and DL methods generally outperformed non-DL baselines.
- Evidence map: 64 studies were included, covering 17 tabular static, 24 tabular temporal, seven genetic or genomic, five image, five signal, and seven multi-modal studies.These counts are not mutually exclusive because one study could involve multiple data types.
- Imputation strategies: 70% (45/64) of studies adopted the separated imputation strategy.
- Imputation strategies: Integrated imputation was popular for tabular temporal and multi-modal data but less used for tabular static, genetic and genomic, image, and signal data.
- Comparisons: Most studies comparing DL with non-DL methods reported superior DL-based approaches in imputation accuracy or downstream prediction performance.Common baselines included simple imputation, RF-based methods, MICE, and k-NN.
- Tabular static data: MLP-oriented models dominated tabular static imputation, with six direct MLP models and five autoencoders using MLP encoders and decoders.
- Tabular temporal data: RNN-based (46%, 11/24) and AE-based (38%, 9/24) methods were the most common approaches for tabular temporal data.Separated imputation was more often paired with AE-based methods than RNN-based methods: 78% (7/9) versus 27% (3/11).
4. Discussion
The discussion links imputation-model design to healthcare data type and missingness structure, emphasizing integrated strategies for complex data while identifying practical and fairness-related boundaries.
- Review contribution: The review organizes deep learning imputation approaches by data type, model backbone, imputation strategy, and downstream-task integration.It considers tabular static, tabular temporal, genetic and genomic, image, signal, and multi-modal data.
- Integrated strategy: The integrated strategy is especially prevalent for tabular temporal and multi-modal data because it accommodates corresponding missing patterns and uncertain overall imputation accuracy.High missing rates limit available ground truth, while multiple data types make a single overall accuracy definition difficult.
- Data-type-specific backbones: MLP-based models and autoencoders suit tabular static data, whereas RNN-based models such as GRU-D capture temporal patterns under informative missingness.Tabular temporal data can involve common informative missingness and high missing rates, complicating characterization of time dynamics.
- Data-type-specific backbones: GAN-based frameworks are commonly used for image imputation, with labels, related images, CNNs, or Transformers supplying additional spatial or cross-image information.The cited discussion describes image translation and attention-based architectures as ways to address image-imputation requirements.
- Integrated strategy: Integrated imputation jointly trains imputation and downstream tasks, allowing their interaction to reduce imputation bias and provide prior information for downstream modeling.The strategy also avoids selecting imputation and downstream models independently, which can be impractical when many combinations must be tested.
- Boundaries and future needs: Integrated approaches remain constrained by additional model complexity, while healthcare applications also require attention to portability, social bias, and fairness.The review notes that most studies addressed statistical bias rather than social bias or discrimination, which can affect subsequent analysis and decision-making.
- Boundaries and future needs: The review’s scope was limited mainly to clinical and translational research and emphasized data types and strategies rather than topic-specific analysis.The authors suggest that other fields and deeper examination of individual research topics may provide additional insights.
5. Conclusions
The review systematically evaluated DL-based missing-value imputation across healthcare data types, summarizing mechanisms, suitable models, and the associated integrated strategy. It identifies portability and fairness as priorities for future research.
- The review fills a gap by systematically evaluating DL-based missing-value imputation methods across varied healthcare data types.
- It summarizes missingness mechanisms and appropriate imputation models for each healthcare data type.
- The review identifies advantages of DL-based imputation models and their corresponding integrated strategy.
- Future research may focus on the portability and fairness of DL-based imputation models in healthcare applications.
Supplementary
The supplementary material presents database-specific search strategies for identifying healthcare studies involving machine learning, missing values, and imputation.
- MEDLINE, Web of Science, Embase, CINAHL, and Scopus searches combine machine-learning terms with missing-data, imputation, and healthcare terms.