Missingness-Aware Prior-Data Fitted Network for Irregular Clinical Time-Series
Authors: Hadi Mehdizavareh, Gabriele Santangelo, Giovanna Nicora, Pietro Bosoni, Simon Lebech Cichosz, Arianna Dagliati, Arijit Khan, Riccardo Bellazzi
Organizations: Aalborg University, Aalborg, Denmark · University of Pavia, Pavia, Italy · Bowling Green State University, Bowling Green, Ohio, USA
Clinical time series are irregularly observed with laboratory measurements recorded at uneven intervals and observation patterns often reflecting patient state, clinical concern, and site-specific practice. Consequently, missingness is not merely a nuisance factor but an informative signal for prediction. We propose a Missingness-Aware Prior-Data Fitted Network (MA-PFN) for irregular clinical time-series prediction. MA-PFN first learns a diffusion-based prior over joint laboratory value and observation-mask trajectories from source-domain electronic health records, capturing both physiological dynamics and measurement processes. Synthetic trajectories sampled from this prior are then transformed into context-query tasks to train a PFN-style conditional predictor. Once trained, the model performs prediction for new admissions in a single forward pass, requiring neither retraining nor parameter adaptation on target datasets. The framework naturally supports both imputation and forecasting through different context-query constructions. We conduct internal validation on a held-out MIMIC-III test set and external validation on eICU, evaluating both imputation and forecasting and assessing cross-domain generalization under shifts in patient populations, clinical practices, and observation mechanisms. MA-PFN achieves the best performance across all evaluated settings, with Overall/Macro nRMSD ranging from (0.193/0.186) for internal imputation to (0.297/0.297) for external forecasting, without target-domain adaptation. These results demonstrate that learning missingness-aware task priors enables accurate and transferable clinical time-series prediction under informative missingness.
Figures & tables
Figure 1: Irregular clinical laboratory observations and their value-mask representation. (a) Three laboratory variables measured asynchronously over time. (b) Alignment to a common time grid produces a sparse value-mask matrix, where missing entries reflect unobserved measurements and may contain information about the underlying clinical observation process.
Figure 2: Overview of the proposed Missingness-Aware Prior-Data Fitted Network (MA-PFN) for irregular clinical time-series prediction.
Method
MIMIC-III
eICU-Imp.
eICU-Fcst.
Mean
0.300/0.300
0.306/0.306
0.349/0.348
LI
0.243/0.241
0.237/0.236
—
LOCF
—
—
0.320/0.320
XGBoost
0.363/0.359
0.406/0.398
0.385/0.377
GRU-D
0.261/0.248
0.292/0.274
0.346/0.341
GraFITi
0.209/0.205
0.237/0.233
0.312/0.310
Table 1: Internal and external evaluation measured by Overall/Macro nRMSD. Imp. and Fcst. denote imputation and forecasting, respectively. LI is used only for imputation and LOCF only for forecasting. Best results are shown in bold and second-best results are underlined . — indicates a task for which the baseline is not applicable.
Figure 3: External forecasting performance on eICU across the seven-day prediction horizon. For each day, nRMSD is computed separately for each laboratory over its observed query targets and macro-averaged across the 13 laboratories.
MIMIC-III Imputation
eICU Imputation
eICU Forecasting
Variant
Overall ↓
Macro ↓
Overall ↓
Macro ↓
Overall ↓
Macro ↓
Direct predictor
0.200
0.198
0.235
0.230
0.325
0.325
PFN (decoupled prior)
0.238
0.237
0.242
0.240
0.309
0.305
MA-PFN
0.193
0.186
0.204
0.198
0.297
0.297
Table 2: Ablation of prior-data fitting and missingness-aware prior learning. The direct predictor is trained only on real MIMIC-III context–query tasks and does not use prior-data fitting. The decoupled-prior PFN retains synthetic prior-data fitting but learns laboratory values and observation masks independently, removing their joint dependence. MA-PFN uses the full learned joint value–mask prior. Lower nRMSD is better.
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
Symbol
Meaning
i
Admission index
L
Number of grid time steps
D
Number of laboratory variables
τ=(τ1,…,τL)
Fixed clinical time grid
X∈RL×D
Normalized laboratory-value matrix
M∈{0,1}L×D
Observation mask, where Mℓd=1 means observed
Appendix
Table 3: Main notation used in the paper and appendix.
Diffusion prior
PFN predictor
Hyperparameter
Selected value
Hyperparameter
Selected value
Denoiser
3-layer BiLSTM
Model dimension d
256
Hidden dimension
256
2D Transformer blocks
4
Diffusion-step embedding
64
Lab-attention heads
4
Diffusion steps
1,000
Time-attention heads
4
Beta schedule
Cosine
Feed-forward dimension
512
Appendix
Table 4: Selected architecture and optimization hyperparameters for the two components of MA-PFN: the missingness-aware diffusion prior and PFN predictor.
Hyperparameter
Search space
Diffusion prior
Loss weight λ
{10−6,10−5,10−4,10−3,10−2,10−1,1}
PFN predictor
Training steps
{50k,100k,200k}
Model dimension ( d )
{64,128,256,512}
Transformer blocks
{4,8,12}
Appendix
Table 5: Hyperparameter search spaces used for MA-PFN model selection. Random search was used for the PFN predictor, while the diffusion prior retained the TimeDiff configuration except for the missingness-loss weight.
Method
Tasks
Representation
Principal hyperparameters and training settings
Mean
Imp./Fcst.
Laboratory identity
Per-laboratory MIMIC-III training mean; no trainable parameters.
LI
Imp.
Per-laboratory observations and times
Piecewise-linear interpolation; nearest-endpoint extrapolation; training-mean fallback when no context value exists.
LOCF
Fcst.
Per-laboratory context history
Most recent pre-query observation; MIMIC-III training-mean fallback; no trainable parameters.
XGBoost
Imp./Fcst.
Query-conditioned grid features
Thirteen independent regressors; squared-error objective; at least two observed training values per laboratory; 252-step imputation windows with stride 128; forecasting context and horizon of 84 and 42 steps, respectively.
GRU-D
Imp./Fcst.
Values, masks, and forward/backward elapsed times
Bidirectional GRU-D; one recurrent layer; recurrent dropout 0.3; prediction-head dropout 0.5; empirical-mean decay; MSE loss; 252-step imputation windows with stride 128; forecasting context and horizon of 84 and 42 steps, respectively.
GraFITi
Imp./Fcst.
Irregular time–variable–value events
Latent dimension 128; one graph-attention layer; four attention heads; AdamW optimizer; learning rate 10−3 ; batch size 64; MSE loss; maximum 200 epochs; learning-rate reduction factor 0.5 with patience 10 and minimum learning rate 10−5 ; early-stopping patience 30.
Appendix
Table 6: Baseline applicability, representations, and principal hyperparameters. Imp. and Fcst. denote imputation and forecasting, respectively. Unless otherwise stated, parameters not listed in the table retained their implementation defaults. All learned baselines were selected using only MIMIC-III training data and were applied to eICU without target-domain adaptation.
Predictor variant
Overall ↓
Macro ↓
Δ Overall
Δ Macro
Temporal inputs
Absolute time + delta time (default)
0.204
0.198
—
—
w/o delta time
0.210
0.203
+0.006
+0.005
w/o absolute time
0.205
0.199
+0.001
+0.001
w/o absolute and delta time
0.219
0.213
+0.015
+0.015
Missingness-aware auxiliary features
Appendix
Table 7: Extended ablations of the PFN predictor on external eICU imputation. All results are Overall/Macro nRMSD. Each group changes only the indicated component relative to the selected configuration in Table 4 . The Δ ) columns report absolute changes relative to the corresponding default, with positive values indicating degradation.
Hyperparameter setting
Value
Overall ↓
Macro ↓
Model dimension (d)
64
0.213
0.207
128
0.206
0.199
256†
0.204
0.198
Number of 2D Transformer blocks
(4†)
0.204
0.198
Appendix
Table 8: Sensitivity of the PFN predictor to model capacity and optimization hyperparameters on external eICU imputation. Results are reported as Overall/Macro nRMSD. The symbol ( † ) marks the selected configuration used in the main experiments. Each experiment changes only the listed hyperparameter while retaining the remaining settings from Table 4 .
Figure 4: Comparison of real and synthetic marginal laboratory-value distributions. For each of the 13 laboratory variables, all observed values are pooled across admissions and time steps after excluding unobserved entries using the corresponding masks. Normalized histograms and kernel density estimate (KDE) curves are shown using shared bin edges for the real and synthetic samples. The generated samples reproduce the central tendency and overall distributional shape of most laboratory variables, with the largest differences occurring in highly skewed or heavy-tailed variables.
Figure 5: UMAP visualization of real and synthetic clinical trajectories. The embedding uses the first week of data, corresponding to 42 four-hour time steps, from randomly selected subsets of 500 real and 500 synthetic trajectories. Each trajectory is represented using all 26 laboratory-value and mask channels, and both groups are jointly embedded in the same two-dimensional space. The substantial overlap and intermixing indicate that the generated trajectories reproduce much of the joint value–missingness structure of the real data, although modest differences in density remain in some peripheral regions.
Method
Offline component
Training time (h)
Mean
None
–
LI
None
–
XGBoost
Training 13 laboratory-specific regressors
0.31±0.05
GRU-D
End-to-end training of the bidirectional recurrent model
6.41±0.23
GraFITi
End-to-end training of the graph-based irregular time-series model
0.73±0.07
CSDI
Training the conditional score-based diffusion model
2.95±0.18
Appendix
Table 9: Offline training time for the imputation methods. Values denote wall-clock time in hours and are reported as mean ± standard deviation over 10 independent runs of the final configurations used to obtain the reported prediction results. For MA-PFN, training the diffusion prior, generating 256,000 synthetic trajectories, and prior-data fitting of the PFN predictor are reported separately. The total MA-PFN time is obtained by summing the three components within each run and subsequently computing the mean and standard deviation across the resulting run-level totals. Dataset extraction, preprocessing, and disk input/output times are excluded. Mean and linear interpolation (LI) require no offline training. TabPFN uses an externally pretrained checkpoint; therefore, its original pretraining cost was not measured in this study.
Figure 6: Inference latency for imputation on the eICU cohort. Bars show the mean end-to-end inference time per admission, and error bars show one standard deviation across admissions. The horizontal axis uses a logarithmic scale. Measurements use batch size one after model warm-up and exclude model loading, dataset loading, and disk input/output, while including context–query construction, required sliding-window processing, and prediction aggregation. All timings use the final configurations employed for the reported accuracy results. CSDI latency includes 50 conditional samples, whereas MA-PFN inference consists only of a PFN forward pass and does not execute the diffusion prior.
Laboratory-Specific Error
Aggregate nRMSD
Method
Metric
Cl
K
HCO 3
Na
Hct
Hgb
MCV
Plt
WBC
RDW
BUN
Cr
Glu
Overall
Macro
No. Queries
8,267
8,267
8,267
8,267
8,267
8,267
8,267
8,267
8,267
8,267
8,267
8,267
8,267
107,471
—
Mean
MAE
3.199
0.390
2.657
2.610
2.645
0.881
1.637
66.69
2.946
0.570
8.427
0.330
31.14
—
—
RMSE
4.225
0.543
3.453
3.488
3.544
1.173
2.292
100.2
4.843
0.876
13.32
0.641
54.30
—
—
nRMSD
0.293
0.280
0.298
0.290
0.295
0.295
0.311
0.319
0.301
0.315
0.316
0.309
0.281
0.300
0.300
LI
MAE
1.912
0.361
1.658
1.782
1.917
0.626
1.186
25.62
1.770
0.269
2.667
0.158
31.67
—
—
Appendix
Table 10: Per-laboratory results for internal imputation on the MIMIC-III test set. “No. Queries” denotes the number of evaluated observed query points for each laboratory, with Overall reporting the total number of queries across all 13 laboratories. MAE and RMSE are reported in the original laboratory units; nRMSD is range-normalized. Overall nRMSD pools all evaluated queries, whereas Macro nRMSD gives equal weight to each laboratory. Lower values are better. Best results are shown in bold and second-best results are underlined .
Laboratory-Specific Error
Aggregate nRMSD
Method
Metric
Cl
K
HCO 3
Na
Hct
Hgb
MCV
Plt
WBC
RDW
BUN
Cr
Glu
Overall
Macro
No. Queries
21,423
21,423
21,423
21,423
21,423
21,423
21,423
21,423
21,423
21,423
21,423
21,423
21,423
278,499
—
Mean
MAE
3.360
0.377
2.660
2.817
2.756
0.930
1.474
56.41
3.025
0.608
8.133
0.356
29.21
—
—
RMSE
4.411
0.498
3.483
3.726
3.742
1.271
2.070
81.92
4.357
0.950
11.73
0.651
44.08
—
—
nRMSD
0.308
0.287
0.307
0.300
0.294
0.296
0.316
0.324
0.305
0.321
0.322
0.310
0.287
0.306
0.306
LI
MAE
1.897
0.310
1.689
1.735
1.817
0.605
0.889
21.37
1.848
0.263
3.382
0.185
26.44
—
—
Appendix
Table 11: Per-laboratory results for external imputation on eICU. All models are evaluated without target-domain training or adaptation. “No. Queries” denotes the number of evaluated observed query points for each laboratory, with Overall reporting the total number of queries across all 13 laboratories. MAE and RMSE are reported in the original laboratory units; lower values are better for all metrics. Best results are shown in bold and second-best results are underlined .
Laboratory-Specific Error
Aggregate nRMSD
Method
Metric
Cl
K
HCO 3
Na
Hct
Hgb
MCV
Plt
WBC
RDW
BUN
Cr
Glu
Overall
Macro
No. Queries
46,575
53,394
46,445
50,257
45,293
46,429
42,142
42,424
41,718
41,961
45,373
46,895
48,233
597,139
—
Mean
MAE
4.271
0.432
3.154
3.754
3.286
1.121
2.274
101.1
3.928
1.157
11.95
0.521
31.35
—
—
RMSE
5.500
0.572
4.094
4.928
4.290
1.467
3.128
144.7
5.452
1.734
16.71
0.895
44.74
—
—
nRMSD
0.346
0.293
0.333
0.332
0.312
0.310
0.382
0.445
0.336
0.428
0.380
0.348
0.283
0.349
0.348
LOCF
MAE
3.558
0.457
2.775
3.309
2.712
0.902
1.657
59.12
3.329
0.699
9.33
0.391
34.95
—
—
Appendix
Table 12: Per-laboratory results for external forecasting on eICU. The first two weeks are used as context and observed laboratory measurements in the subsequent week are prediction targets. “No. Queries” denotes the number of evaluated observed query points for each laboratory, with Overall reporting the total number of queries across all 13 laboratories. MAE and RMSE are reported in the original laboratory units; lower values are better for all metrics. Best results are shown in bold and second-best results are underlined .
Laboratory tests in electronic health records are collected irregularly, and the absence of a test order can be as informative as the measurement itself. Such missingness reflects clinicians' decisions and patient physiology, making it important to model it directly rather than treat it as a preprocessing artifact. Here we present a diffusion-based approach for generating clinical time series that jointly models laboratory values and their observation patterns using the public Data Analytics Challenge on Missing Data Imputation (DACMI) benchmark derived from MIMIC-III. To preserve realistic sampling, we align chart times into 4-hour intervals and segment admissions into 7-day windows, producing trajectories that pair each lab value with a corresponding observation indicator. Standard transformations and normalization are applied to stabilize training. Our method extends the TimeDiff framework to learn continuous lab values and discrete missingness patterns through complementary diffusion objectives. Experiments show that the generated data closely match real patient trajectories across individual lab distributions and joint value-missingness embeddings, demonstrating that diffusion models can capture clinically meaningful dependencies between patient physiology and clinicians' testing behavior under MNAR-like (missing-not-at-random) missingness. These preliminary results indicate that our model can serve as an initial component toward developing clinical foundation models. By producing synthetic priors that preserve key physiology-missingness relationships, this work motivates the subsequent training of Prior-Data Fitted Networks capable of leveraging informative missingness, which we will investigate in the extended work.
Hadi Mehdizavareh, Gabriele Santangelo, Giovanna Nicora +4
Aalborg University, Aalborg, Denmark · University of Pavia, Pavia, Italy · Bowling Green State University, Bowling Green, Ohio, USA
In healthcare, multimodal time series tasks often operate on incomplete observations in practice, for example when ECG segments are lost because electrodes detach or an entire respiratory channel is unavailable during overnight monitoring. Such missingness typically appears in two structurally distinct patterns: within-modality missing, where values are absent within an otherwise observed modality, and modality-level missing, where an entire modality is unavailable. Existing methods typically represent unobserved data implicitly through masks or missing embeddings, without learning instance-specific missing information, and most are designed for only one missingness pattern. A natural approach is to explicitly estimate the missing data; however, existing imputation methods treat missingness uniformly despite their different structural priors, and the imputation process is often isolated from downstream tasks, preventing downstream tasks from guiding imputation toward more informative representations. To address these limitations, we present PAMF, a multimodal time-series framework that explicitly handles different missingness patterns while coupling imputation with downstream prediction through prior-aware flow matching and weight sharing. Specifically, the method initializes the flow-matching source state with type-specific priors to distinguish two missing types. It further connects imputation and classification through architecturally matched encoders with weight sharing, transferring task-relevant representations into the imputation process. Experiments on multiple multimodal healthcare time-series benchmarks show that the proposed method achieves the strongest overall downstream performance across diverse datasets and missing settings compared with existing baselines.
Ziwen Kan, Wugeng Zheng, Tianlong Chen +1
Department of Computer Science, University of Central Florida · Department of Computer Science, University of North Carolina at Chapel Hill
Clinical time series prediction in intensive care units remains challenging due to heterogeneous physiological variables and informative missingness. The presence or absence of a measurement can reflect clinical decisions and patient severity, and thus missingness can serve as a predictive signal rather than a simple data artifact. This work presents CISM, a Channel-Independent Spectrogram framework with a Missingness stream for clinical multivariate time series prediction. CISM converts each clinical variable into a variable-wise time-frequency spectrogram, preserves variable identity through variable-aligned encoding, and aligns an explicit missingness stream with the spectrogram representation. Experiments on an in-hospital mortality task derived from MIMIC-IV show that CISM achieves the highest mean AUROC (0.7225), AUPRC (0.3308), and F1 (0.3808) among the compared time series, missingness-aware, vision, and time-frequency baselines. Ablation studies further show that observation patterns provide a meaningful informative signal. Pixel-level mask injection improves performance over plain spectrogram inputs and recovers much of this predictive value. The aligned missingness stream contributes a further, complementary gain in both AUROC and AUPRC. These results highlight the importance of modeling observation patterns as structured signals in clinical time series prediction.
Soyeon Park, Charmgil Hong
Handong Global University Pohang, Republic of Korea