Handling missing values in healthcare data: A systematic review of deep learning-based imputation techniques

1. Introduction

• New healthcare data formats are more susceptible to missingness than traditional tabular clinical data, and simple strategies may not suit diverse data types and architectures.The examples include monitoring-device and sequencing data, and the range spans static and temporal, tabular, imaging, and sequencing data.Source: §1 · p. 4
• Missing values can reduce predictive power and bias downstream decisions, a particular concern in high-stakes healthcare settings.The introduction notes that simple complete-case analysis or mean, median, or mode imputation may address static tabular data but may not suit other data types.Source: §1 · p. 4

imputation by chained equations (MICE)[6], random forest (RF)-based models like

• Conventional imputation methods may be limited to certain missing-data categories, and MICE generally assumes data are missing at random [6].These non-deep-learning strategies may also have low accuracy on complex healthcare data, especially when the missingness mechanism is uncertain [8, 9].Source: imputation by chained equations… · p. 4
• Deep learning methods have demonstrated improved imputation accuracy [10, 11] and can be tailored to complex missing patterns and data structures.The introduction gives time-series sequential structure and image spatial patterns as examples [12, 13].Source: imputation by chained equations… · p. 5
• Existing reviews either focused on non-DL methods [14, 15] or treated neural networks as a single method [16, 17], limiting guidance for researchers applying DL models to their own data.The authors knew of no systematic review of DL-based missing-value imputation.Source: imputation by chained equations… · p. 5
• The review presents an evidence map examining deep-learning model use by data type to guide researchers managing missing values across healthcare disciplines.The evidence-map analysis is identified as [18].Source: imputation by chained equations… · p. 5

2.1 Search Strategies

• Five databases were searched using terms covering missing values, imputation, machine learning, deep learning, and healthcare.The databases were MEDLINE, Web of Science, Embase, CINAHL, and Scopus.Source: §2.1 · p. 5

2.2 Exclusion Criteria

• Studies were assessed under PRISMA guidelines[19] and excluded if outside the medical or clinical domain, not DL-based, not a research article, or not in English.Source: §2.2 · p. 6

2.3 Selection Procedure and Data Extraction

• Two reviewers independently screened studies and extracted data, resolving disagreements with a third reviewer; extracted fields covered data type, model backbone, imputation strategy, and non-DL comparisons.Source: §2.3 · p. 6

2.4 Data Analysis

• The evidence map classified data as tabular static, tabular temporal, genetic/genomic, image, signal, or multi-modal; temporal data include a time factor, and multi-modal tasks use multiple data types.Source: §2.4 · p. 6

which includes vanilla autoencoder, denoising autoencoder (DAE), and variational autoencoder

• Model backbones were grouped as MLP, RNN, autoencoder, GAN, hybrids of these architectures, or other models such as SOM.Source: §2.4 · p. 7, which includes vanilla autoencoder,… · p. 7

2.4 Data Analysis

• Imputation strategies were categorized as separated from downstream tasks or integrated with them concurrently; backbone and data-type cross-tabulation formed the evidence map.A bar plot displayed the distribution of strategies.Source: which includes vanilla autoencoder,… · p. 7

3. Results

• The review counted 17 static-tabular, 24 temporal-tabular, seven genetic/genomic, five image, five signal, and seven multimodal studies; MLP for static and vanilla autoencoder for temporal data were most common (9.4%, 6/64 among cross-tabulation cells).Data-type counts are not mutually exclusive because a study may involve more than one data type.Source: §3 · p. 7
• 70% (45/64) of studies used separated imputation; 35 studies examined missingness, including 18 that investigated MNAR.The integrated strategy was more common for tabular temporal and multi-modal data, less used for tabular static and genetic/genomic data, and rare for image or signal data.Source: §3 · p. 8

3.1 Tabular Static Data

• Most tabular-static studies demonstrated superior imputation accuracy for DL-based approaches; simple imputation, RF-based methods, MICE, and k-NN were common comparison baselines.Among 17 studies, 76% (13/17) provided non-DL baselines; the most common were simple imputation (41%, 7/17), RF-based methods (35%, 6/17), MICE (36%, 6/17), and k-NN (29%, 5/17).Source: §3.1 · p. 8
• Tabular-static methods were dominated by MLP-oriented models, and DL methods could accommodate mixed variable types with suitable encoding and activation.Six studies used MLPs directly, five used autoencoders with MLP encoders and decoders, and three GAN studies used MLP generators and discriminators.Source: §3.1 · pp. 8–9

3.2 Tabular Temporal Data

• In tabular temporal imputation studies, RNN-based methods were used in 46% (11/24) and AE-based methods in 38% (9/24); separated imputation was more frequent with AE-based methods (78%, 7/9) than RNN-based methods (27%, 3/11).Source: §3.2 · p. 9
• In temporal-data studies, RNNs could capture time-dynamics-related missing patterns, while GAN approaches were described as helping avoid error propagation into downstream tasks.Two studies used GANs alone or combined with RNNs for this purpose.Source: §3.2 · pp. 9–10

3.3 Genetic and Genomic Data

• Seven genetic/genomic studies used AE-based models in four studies and MLPs in three; the review noted potential for biological knowledge and interpretable imputation.The AE/VAE architecture was described as allowing more interpretable imputation procedures.Source: §3.3 · p. 10

3.4 Image Data

• All five image-imputation studies used GAN frameworks, but only one established non-DL baselines; conditional GANs could address missing pixels or an entirely missing image.The sole study with baselines compared against mean imputation and interpolation.Source: §3.4 · p. 11
• For MRI/PET consistency imputation, Cycle-GAN and Colla-GAN used two discriminators, paralleling Pan et al.'s two-generator, two-discriminator design linking the modalities [12, 68, 69].The discriminators detected synthetic MRI and PET, respectively.Source: §3.4 · p. 11

3.5 Signal Data

• Signal data typically have high sample rates and are susceptible to noise; during movement, intermittent disconnections, body movement, or firmware malfunction can cause large blocks of concurrently missing features [74].Source: §3.5 · p. 12
• Simple mean/median imputation was the common baseline in 80% (4/5) of signal studies; three used customized autoencoders and two used RNNs [74, 76].Customizations included treating missingness as noise and using adversarial autoencoders.Source: §3.5 · p. 12
• For fMRI signal imputation, spatial information first filled missing values and a GRU regularized the time domain, using spatiotemporal structure to assist imputation [76].The review also noted that noise artifacts make missing values common in fMRI.Source: §3.5 · p. 12

3.6 Multi-modal Data

• Seven studies examined multi-modal imputation, where modality fusion was essential; four used autoencoders linking encoders and decoders across modalities.Other designs linked modality-specific MLPs or enabled iterative communication between modality-specific layers.Source: §3.6 · pp. 12–13
• Combining graph convolution with LSTM was effective for multi-modal imputation framed as geometric-matrix completion.The framework used self-attention to aggregate recurrent graph-convolutional models.Source: §3.6 · p. 13

Discussion

• Data-type-oriented deep-learning models can account for corresponding missing patterns, improving imputation; integrated strategies can also enhance imputation and downstream analysis.The review reports that integrated strategies were relatively prevalent for tabular temporal and multi-modal data, consistent with the characteristics of those data types.Source: §4 · pp. 13, 15
• MLPs and MLP-based autoencoders suit tabular static data, while GRU-D variants address temporal patterns amid informative missingness and high missing rates.For genetic and genomic data, MLPs or autoencoders can effectively use biological knowledge.Source: §4 · p. 13
• GAN-based models are commonly used for image imputation; labels or relevant images can improve performance, and CNN-to-Transformer modules can address spatial information.Hybrid models can capture interactions among missingness patterns; Vivar et al.’s LSTM and graphical-convolution integration advanced imputation and classification in computer-aided diagnosis [81].Source: §4 · p. 14
• The review identifies a vacancy in medical-text imputation and urges attention to fairness, since social bias or discrimination can affect subsequent analysis and lead to unfair decision-making.Well-packaged deep-learning models specified by data type are of practical value.Source: §4 · pp. 14, 16
• Integrated training can reduce imputation bias, provide prior information for downstream modeling, and avoid selecting imputation and downstream models separately.It can be practical when ground truth is limited or overall accuracy is difficult to define, and it avoids testing time-consuming combinations; however, added model complexity can restrict adoption.Source: §4 · p. 15
• For complex healthcare data, deep-learning models may avoid restrictive statistical assumptions and extensive feature engineering, but non-deep-learning methods can perform well on carefully engineered tabular data.The review notes non-deep-learning limitations including high dimensionality and low time efficiency, while pretrained deep-learning models may reduce evaluation costs when applied to new samples [61].Source: §4 · pp. 15–16
• The review was limited to clinical and translational research and focused on data types and imputation strategies rather than specific research topics.The authors suggest examining other fields and healthcare applications by research topic in future work.Source: §4 · p. 16

• The review addresses a gap by systematically evaluating DL-based imputation across healthcare data types and summarizing missingness mechanisms and appropriate models for each type.Source: §5 · p. 17
• The study identifies advantages of DL-based imputation models and their corresponding integrated strategy.Source: §5 · p. 17
• Future research may focus on the portability and fairness of DL-based imputation models in healthcare applications.Source: §5 · p. 17

eTable 1. Search Strategies

• MEDLINE’s strategy combines machine-learning, deep-learning, AI, or data-mining terms with missing-data, imputation, and medical or healthcare terms.Source: eTable 1. Search Strategies · p. 34
• Web of Science’s strategy combines machine-learning, deep-learning, AI, or data-mining terms with missing-data and clinical or healthcare terms.Source: eTable 1. Search Strategies · p. 34
• Embase’s strategy combines machine-learning, deep-learning, AI, or data-mining terms with missing-data, imputation, and medical or healthcare terms.Source: eTable 1. Search Strategies · p. 34
• CINAHL’s strategy combines machine-learning, deep-learning, AI, or data-mining terms with missing-data, imputation, and medical or healthcare terms.Source: eTable 1. Search Strategies · p. 34
• Scopus’s strategy combines machine-learning, deep-learning, AI, or data-mining terms with missing-data, imputation, and medical or healthcare terms.Source: eTable 1. Search Strategies · p. 34