As artificial intelligence systems move toward clinical deployment, ensuring reliable prediction behavior is fundamental for safety-critical decision-making tasks. One proposed safeguard is selective prediction, where models can defer uncertain predictions to human experts for review. In this work, we empirically evaluate the reliability of uncertainty-based selective prediction in multilabel clinical condition classification using multimodal ICU data. Across a range of state-of-the-art unimodal and multimodal models, we find that selective prediction can substantially degrade performance despite strong standard evaluation metrics. This failure is driven by severe class-dependent miscalibration, whereby models assign high uncertainty to correct predictions and low uncertainty to incorrect ones, particularly for underrepresented clinical conditions. Our results show that commonly used aggregate metrics can obscure these effects, limiting their ability to assess selective prediction behavior in this setting. Taken together, our findings characterize a task-specific failure mode of selective prediction in multimodal clinical condition classification and highlight the need for calibration-aware evaluation to provide strong guarantees of safety and robustness in clinical AI.
Figures & tables
Figure 1: Selective Prediction Can Serve as a Fail-Safe Mechanism in Safety-Critical Healthcare Settings. a) A model outputs prediction confidence/uncertainty scores; low-confidence cases are rejected and referred to an expert, yielding coverage-controlled evaluation and diagnosis. b) Miscalibration causes predictions that are mostly correct to have high uncertainty (low confidence) and incorrect predictions to have high confidence. Theoretical selective prediction curves show the difference between a well-calibrated model, random selection, and a severely miscalibrated model, showing improving, flat, and degrading performance with increasing coverage. Clinically, overconfidence risks missed or delayed diagnoses, while underconfidence triggers unnecessary rejections and shifts excess workload to clinicians.
Model
AUROC
AUPRC
Selective AUROC
Selective AUPRC
ECE↓
EHR (LSTM) ( Hochreiter and Schmidhuber, 1997 )
0.681 ± 0.007
0.371 ± 0.010
0.625 ± 0.115
0.292 ± 0.091
1.42 ± 0.335
CXR (ResNet-34) ( He et al., 2015 )
0.654 ± 0.013
0.350 ± 0.012
0.643 ± 0.095
0.342 ± 0.077
9.41 ± 0.826
DrFuse ( Yao et al., 2024 )
0.726 ± 0.007
0.418 ± 0.010
0.677 ± 0.139
0.352 ± 0.116
2.54 ± 1.15
MeTra ( Khader et al., 2023 )
0.707 ± 0.010
0.399 ± 0.015
0.661 ± 0.125
0.333 ± 0.104
2.32 ± 0.83
MedFuse ( Hayat et al., 2022 )
0.739 ± 0.005
0.449 ± 0.008
0.705 ± 0.136
0.396 ± 0.112
2.08 ± 0.725
MedFuse + Temp. Scaling ( Guo et al., 2017 )
0.739 ± 0.005
0.449 ± 0.008
0.705 ± 0.135
0.398 ± 0.110
1.97 ± 0.60
Table 1: Performance Across Conditions. Aggregate metrics suggest that multimodal learning (MedFuse) outperforms unimodal baselines. However, these results mask substantial condition-specific variability. Particularly, standard calibration measures (ECE) are deceptively low, concealing severe and consistent miscalibration across individual clinical conditions. This condition-specific variability, motivates the class-level and selective prediction analysis presented in our study. Other multimodal architectures (DrFuse, MeTra) achieve similar aggregate performance, yet, we show that they exhibit comparable calibration failures despite more complex fusion mechanisms. We report improvement gains over EHR given its widespread use as the foundation for most clinical prediction models. ( Best , Second Best ).
Figure 2: How Does Calibration Influence Selective Prediction Behavior? (a) Selective AUROC, selective AUPRC, and class-stratified ECE across 25 clinical conditions for MedFuse as low-confidence predictions are rejected. (b) Relationship between positive-class calibration error (ECE c=1 AUC) and selective AUROC/AUPRC across conditions and random seeds, showing worse selective performance with higher ECE c=1 (Spearman’s rank, p<0.05 . Shaded region: variance across seeds.)
Figure 3: Are Class-dependent Calibration Effects Consistent Across Multimodal Architectures? Across MedFuse, DrFuse, and MeTRA, higher calibration error in the underrepresented class (i.e., positive class in this task) is consistently associated with lower selective AUROC and AUPRC across conditions. For all architectures and metrics, the null hypothesis of no association between stratified ECE and selective AUC is rejected, indicating that architectural complexity alone does not mitigate calibration-driven selective prediction failures. (Spearman’s rank H0 , p<0.05 )
Figure 4: Can Loss-Upweighting Mitigate Observed Calibration Failures? Across architectures loss-upweighting reduces calibration error for the underrepresented class, but the negative relationship between stratified ECE and selective AUROC/AUPRC persists across conditions. For all architectures and metrics, the null hypothesis of no association between ECE and selective AUC is rejected (Spearman’s rank, p<0.05 ).
Appendix figures & tables19 assets
Supplementary material from the paper’s appendix.
Appendix
Clinical Condition
Organ
Modality
Train Set Prevalence
CXR
EHR
Positive
Negative
1 Acute and unspecified renal failure
Kidneys
-
✓
32.32% (2,532)
67.68% (5,309)
2 Acute cerebrovascular disease
Brain (Cerebrovascular system)
-
✓
7.99% (625)
92.01% (7,216)
3 Acute myocardial infarction
Heart (Coronary arteries)
-
✓
8.57% (675)
91.43% (7,166)
4 Cardiac dysrhythmias
Heart (Conduction system)
-
✓
37.36% (2,942)
62.64% (4,899)
5 Chronic kidney disease
Kidneys
✓
✓
23.34% (1,824)
76.66% (6,017)
Appendix
Table A.1 : Clinical Conditions and Dataset Characteristics. Overview of the 25 conditions analyzed in this study, including their associated organ/system, primary diagnostic modality (CXR or EHR), and the prevalence of positive and negative samples in the training set.
Clinical Condition
EHR
CXR
MedFuse
AUROC
AUPRC
Selective AUROC
Selective AUPRC
ECE↓
AUROC
AUPRC
Selective AUROC
Selective AUPRC
ECE↓
AUROC
AUPRC
Selective AUROC
Selective AUPRC
ECE↓
1
0.724 ± 0.003
0.564 ± 0.010
0.772 ± 0.100
0.627 ± 0.100
1.62 ± 0.52
0.639 ± 0.009
0.453 ± 0.014
0.684 ± 0.081
0.503 ± 0.075
12.77 ± 0.98
0.761 ± 0.005
0.589 ± 0.008
0.804 ± 0.111
0.634 ± 0.109
2.31 ± 0.97
2
0.873 ± 0.005
0.459 ± 0.010
0.569 ± 0.180
0.085 ± 0.138
1.08 ± 0.14
0.650 ± 0.009
0.139 ± 0.009
0.577 ± 0.121
0.084 ± 0.057
3.99 ± 0.31
0.856 ± 0.010
0.432 ± 0.019
0.560 ± 0.176
0.075 ± 0.116
0.88 ± 0.24
3
0.693 ± 0.014
0.175 ± 0.016
0.617 ± 0.150
0.089 ± 0.070
0.73 ± 0.07
0.640 ± 0.015
0.153 ± 0.017
0.567 ± 0.102
0.092 ± 0.037
4.35 ± 0.42
0.713 ± 0.005
0.209 ± 0.010
0.633 ± 0.207
0.083 ± 0.049
0.76 ± 0.23
4
0.598 ± 0.004
0.462 ± 0.012
0.618 ± 0.069
0.491 ± 0.065
1.15 ± 0.67
0.647 ± 0.009
0.534 ± 0.013
0.683 ± 0.077
0.598 ± 0.077
12.50 ± 0.48
0.694 ± 0.006
0.565 ± 0.010
0.727 ± 0.079
0.610 ± 0.071
2.61 ± 1.65
5
0.696 ± 0.006
0.402 ± 0.012
0.642 ± 0.113
0.244 ± 0.113
2.13 ± 0.45
0.662 ± 0.021
0.378 ± 0.019
0.696 ± 0.085
0.409 ± 0.078
15.37 ± 0.37
0.770 ± 0.003
0.519 ± 0.010
0.787 ± 0.177
0.521 ± 0.169
2.34 ± 2.04
Appendix
Table C.2 : Do Multimodal Models Improve or Degrade Selective Prediction Performance in this Task? Condition-level performance of standard and selective prediction metrics. MedFuse generally improves AUROC, AUPRC and selective AUROC, but calibration gains are very inconsistent compared to EHR. (Dark-bold: p<0.05 , Wilcoxon signed-rank test, 5 seeds; Light-bold: highest mean, not significant)
Clinical Condition
MedFuse
DrFuse
MeTra
AUROC
AUPRC
Selective AUROC
Selective AUPRC
ECE↓
AUROC
AUPRC
Selective AUROC
Selective AUPRC
ECE↓
AUROC
AUPRC
Selective AUROC
Selective AUPRC
ECE↓
1
0.761 ± 0.005
0.589 ± 0.008
0.804 ± 0.111
0.634 ± 0.109
2.31 ± 0.97
0.732 ± 0.004
0.551 ± 0.008
0.767 ± 0.104
0.585 ± 0.113
4.06 ± 1.21
0.722 ± 0.009
0.541 ± 0.011
0.759 ± 0.083
0.573 ± 0.071
2.73 ± 0.39
2
0.856 ± 0.010
0.432 ± 0.019
0.560 ± 0.176
0.075 ± 0.116
0.88 ± 0.24
0.899 ± 0.004
0.460 ± 0.013
0.584 ± 0.272
0.118 ± 0.166
1.06 ± 0.44
0.900 ± 0.004
0.490 ± 0.022
0.602 ± 0.253
0.112 ± 0.172
1.13 ± 0.32
3
0.713 ± 0.005
0.209 ± 0.010
0.633 ± 0.207
0.083 ± 0.049
0.76 ± 0.23
0.697 ± 0.013
0.166 ± 0.010
0.637 ± 0.169
0.090 ± 0.069
2.21 ± 0.58
0.717 ± 0.010
0.197 ± 0.014
0.610 ± 0.158
0.085 ± 0.058
1.98 ± 1.12
4
0.694 ± 0.006
0.565 ± 0.010
0.727 ± 0.079
0.610 ± 0.071
2.61 ± 1.65
0.650 ± 0.012
0.520 ± 0.013
0.692 ± 0.084
0.547 ± 0.089
3.42 ± 1.66
0.628 ± 0.011
0.500 ± 0.008
0.648 ± 0.076
0.529 ± 0.079
3.59 ± 1.65
5
0.770 ± 0.003
0.519 ± 0.010
0.787 ± 0.177
0.521 ± 0.169
2.34 ± 2.04
0.726 ± 0.006
0.448 ± 0.009
0.693 ± 0.127
0.334 ± 0.152
3.59 ± 2.23
0.709 ± 0.017
0.414 ± 0.019
0.668 ± 0.101
0.299 ± 0.105
2.68 ± 1.31
Appendix
Table C.3 : Condition-level Performance of Evaluation Metrics Across Multimodal Architectures. MedFuse, DrFuse, and MeTra achieve broadly comparable discrimination performance across conditions, with MedFuse consistently attaining the highest AUROC/AUPRC values, while all models exhibit elevated calibration error. Shading and bolding are applied within each model to compare standard AUC metrics against their selective AUC counterparts for the same condition. Under correct uncertainty estimation, selective AUROC/AUPRC should be statistically significantly higher than standard metrics, reflecting effective rejection of uncertain predictions. Instead, we observe no consistent improvement across architectures: standard AUC metrics frequently dominate their selective counterparts, indicating that selective prediction fails to provide systematic gains. These results further suggest that increased architectural complexity alone does not improve reliability or resolve calibration-driven failures in this task. (Dark-bold: p<0.05 , Wilcoxon signed-rank test, 5 seeds; Light-bold: highest mean, not significant)
Clinical Condition
MedFuse (Loss-Upweighting)
DrFuse (Loss-Upweighting)
MeTra (Loss-Upweighting)
AUROC
AUPRC
Selective AUROC
Selective AUPRC
ECE↓
AUROC
AUPRC
Selective AUROC
Selective AUPRC
ECE↓
AUROC
AUPRC
Selective AUROC
Selective AUPRC
ECE↓
1
0.761 ± 0.003
0.588 ± 0.008
0.756 ± 0.088
0.672 ± 0.065
3.09 ± 0.86
0.731 ± 0.004
0.547 ± 0.007
0.717 ± 0.090
0.628 ± 0.081
4.58 ± 0.98
0.715 ± 0.008
0.533 ± 0.012
0.721 ± 0.088
0.618 ± 0.077
3.35 ± 1.48
2
0.871 ± 0.006
0.431 ± 0.010
0.872 ± 0.093
0.491 ± 0.069
2.38 ± 0.66
0.895 ± 0.005
0.473 ± 0.010
0.878 ± 0.153
0.547 ± 0.135
4.44 ± 1.50
0.897 ± 0.007
0.478 ± 0.034
0.871 ± 0.111
0.546 ± 0.078
3.79 ± 0.79
3
0.716 ± 0.003
0.205 ± 0.005
0.744 ± 0.167
0.260 ± 0.097
9.58 ± 2.12
0.693 ± 0.007
0.175 ± 0.007
0.687 ± 0.118
0.205 ± 0.073
13.34 ± 6.55
0.699 ± 0.027
0.182 ± 0.012
0.694 ± 0.143
0.205 ± 0.070
10.09 ± 3.99
4
0.692 ± 0.005
0.563 ± 0.007
0.650 ± 0.084
0.631 ± 0.060
2.89 ± 1.26
0.621 ± 0.005
0.492 ± 0.005
0.589 ± 0.073
0.550 ± 0.066
4.58 ± 2.03
0.630 ± 0.009
0.496 ± 0.010
0.595 ± 0.073
0.543 ± 0.068
1.93 ± 0.65
5
0.760 ± 0.006
0.496 ± 0.012
0.805 ± 0.094
0.599 ± 0.113
4.60 ± 1.52
0.718 ± 0.002
0.437 ± 0.006
0.726 ± 0.085
0.522 ± 0.088
5.90 ± 2.72
0.696 ± 0.015
0.402 ± 0.013
0.678 ± 0.095
0.467 ± 0.080
4.98 ± 3.28
Appendix
Table C.5 : Condition-level Performance of Evaluation Metrics Across Multimodal Architectures under Loss-Upweighting. Although the performance of all three models is comparable, MedFuse outperforms both DrFuse and MeTra across different conditions and evaluation metrics, with the downside that all models present high ECE scores and the difference between models is not statistically significant. This further confirms that complex architectures do not necessarily improve predictive performance and do not solve the reliability issues observed in this clinical task. (Dark-bold: p<0.05 , Wilcoxon signed-rank test, 5 seeds; Light-bold: highest mean, not significant)
Clinical Condition
MedFuse (Loss-Upweighting)
DrFuse (Loss-Upweighting)
MeTra (Loss-Upweighting)
ECE↓
ECEc=1↓
ECEc=0↓
ECE↓
ECEc=1↓
ECEc=0↓
ECE↓
ECEc=1↓
ECEc=0↓
1
3.09 ± 0.86
3.81 ± 1.13
5.01 ± 1.96
4.58 ± 0.98
4.57 ± 0.84
6.32 ± 1.96
3.35 ± 1.48
7.19 ± 4.48
7.36 ± 2.53
2
2.38 ± 0.66
10.82 ± 1.46
2.10 ± 0.76
4.44 ± 1.50
9.75 ± 2.09
4.69 ± 1.66
3.79 ± 0.79
5.78 ± 1.66
4.41 ± 0.85
3
9.58 ± 2.12
12.14 ± 1.93
11.33 ± 2.29
13.34 ± 6.55
11.71 ± 2.50
15.62 ± 7.08
10.09 ± 3.99
18.20 ± 13.24
12.53 ± 5.48
4
2.89 ± 1.26
7.52 ± 2.45
5.32 ± 5.21
4.58 ± 2.03
7.31 ± 2.60
11.83 ± 4.69
1.93 ± 0.65
11.11 ± 5.16
8.07 ± 3.52
5
4.60 ± 1.52
5.05 ± 2.19
5.40 ± 2.97
5.90 ± 2.72
5.60 ± 2.70
8.21 ± 4.41
4.98 ± 3.28
12.78 ± 12.05
9.92 ± 7.48
Appendix
Table C.6 : Condition-level Performance of ECE Stratification Across Multimodal Architectures under Loss-Upweighting. Class-stratified calibration metrics under loss upweighting show reduced positive-class ECE for many conditions, but it still remains dominant across architectures. No single multimodal model consistently achieves the lowest ECE across conditions, indicating that loss upweighting mitigates but does not resolve class-dependent calibration errors. (Dark-bold: p<0.05 , Wilcoxon signed-rank test, 5 seeds; Light-bold: highest mean, not significant)
Clinical Condition
EHR
CXR
MedFuse
AECE↓
AECEc=1↓
AECEc=0↓
AECE↓
AECEc=1↓
AECEc=0↓
AECE↓
AECEc=1↓
AECEc=0↓
1
2.32 ± 0.48
57.80 ± 0.36
27.38 ± 0.09
14.39 ± 1.15
56.87 ± 3.22
29.87 ± 3.00
3.22 ± 0.91
54.09 ± 2.26
25.83 ± 2.03
2
1.02 ± 0.21
66.59 ± 0.37
6.23 ± 0.14
5.87 ± 0.47
86.92 ± 2.62
6.84 ± 1.62
1.29 ± 0.26
67.96 ± 1.40
5.40 ± 0.18
3
1.37 ± 0.12
88.27 ± 0.33
8.22 ± 0.06
6.75 ± 1.33
85.30 ± 3.50
8.38 ± 2.50
1.46 ± 0.30
86.93 ± 0.14
8.05 ± 0.25
4
2.37 ± 0.50
60.63 ± 0.13
36.44 ± 0.12
14.13 ± 1.07
52.81 ± 4.49
32.66 ± 4.52
3.93 ± 1.67
55.81 ± 3.90
32.44 ± 3.51
5
2.65 ± 0.37
70.95 ± 0.28
21.08 ± 0.07
17.08 ± 0.70
59.96 ± 1.82
22.65 ± 1.99
3.64 ± 1.62
61.94 ± 4.35
18.95 ± 3.07
Appendix
Table C.9 : Adaptive ECE for Unimodal Models. Aggregate and class-stratified Adaptive Expected Calibration Error (AECE) are reported for unimodal baselines. Positive-class AECE remains substantially higher than negative-class AECE, confirming that minority-class miscalibration persists when calibration error is estimated using adaptive equal-frequency bins. (Dark-bold: p<0.05 , Wilcoxon signed-rank test, 5 seeds; Light-bold: highest mean, not significant)
Clinical Condition
MedFuse
DrFuse
MeTra
AECE↓
AECEc=1↓
AECEc=0↓
AECE↓
AECEc=1↓
AECEc=0↓
AECE↓
AECEc=1↓
AECEc=0↓
1
3.22 ± 0.91
54.09 ± 2.26
25.83 ± 2.03
5.16 ± 1.20
55.57 ± 3.44
26.30 ± 2.71
3.01 ± 0.39
59.40 ± 1.02
26.08 ± 1.02
2
1.29 ± 0.26
67.96 ± 1.40
5.40 ± 0.18
1.92 ± 1.05
60.49 ± 4.83
6.21 ± 1.11
2.02 ± 0.62
59.96 ± 5.55
6.44 ± 1.01
3
1.46 ± 0.30
86.93 ± 0.14
8.05 ± 0.25
2.37 ± 0.35
88.15 ± 2.27
7.83 ± 1.71
2.11 ± 1.13
84.75 ± 2.92
9.34 ± 1.95
4
3.93 ± 1.67
55.81 ± 3.90
32.44 ± 3.51
4.82 ± 2.29
58.41 ± 5.60
34.07 ± 4.56
5.16 ± 1.72
59.83 ± 4.55
34.46 ± 4.44
5
3.64 ± 1.62
61.94 ± 4.35
18.95 ± 3.07
4.64 ± 2.03
66.72 ± 4.71
19.40 ± 3.56
3.36 ± 1.01
70.00 ± 1.94
20.06 ± 1.60
Appendix
Table C.10 : Adaptive ECE for Multimodal Models. Aggregate and class-stratified Adaptive Expected Calibration Error (AECE) are reported for standard multimodal architectures. The positive class exhibits substantially higher AECE than the negative class across MedFuse, DrFuse, and MeTra, reinforcing that multimodal performance gains do not eliminate class-dependent calibration errors. (Dark-bold: p<0.05 , Wilcoxon signed-rank test, 5 seeds; Light-bold: highest mean, not significant)
Clinical Condition
MedFuse (Loss-Upweighting)
DrFuse (Loss-Upweighting)
MeTra (Loss-Upweighting)
AECE↓
AECEc=1↓
AECEc=0↓
AECE↓
AECEc=1↓
AECEc=0↓
AECE↓
AECEc=1↓
AECEc=0↓
1
14.47 ± 1.33
38.49 ± 1.01
39.54 ± 1.49
14.91 ± 1.51
40.19 ± 1.19
40.42 ± 1.72
13.67 ± 3.13
43.47 ± 3.99
40.70 ± 2.76
2
15.85 ± 1.07
28.20 ± 1.19
19.61 ± 1.08
15.46 ± 3.11
26.76 ± 3.93
18.51 ± 3.20
21.47 ± 3.66
22.96 ± 1.76
25.26 ± 3.93
3
30.73 ± 2.15
43.34 ± 2.27
38.31 ± 2.15
35.58 ± 5.53
41.28 ± 5.99
43.24 ± 5.51
31.64 ± 7.15
46.78 ± 8.88
39.66 ± 6.98
4
11.02 ± 2.99
43.65 ± 2.86
44.38 ± 3.15
12.31 ± 1.80
46.40 ± 1.75
48.18 ± 1.84
9.66 ± 2.90
48.86 ± 3.10
45.35 ± 2.94
5
19.10 ± 2.90
39.06 ± 2.45
37.57 ± 3.07
22.74 ± 3.28
41.06 ± 3.14
42.39 ± 3.37
21.17 ± 5.49
45.53 ± 6.87
42.35 ± 5.08
Appendix
Table C.11 : Adaptive ECE for Multimodal Models under Loss-Upweighting. Aggregate and class-stratified Adaptive Expected Calibration Error (AECE) are reported for multimodal architectures trained with loss upweighting. Loss upweighting reduces positive-class AECE relative to standard multimodal models, but it redistributes calibration error toward the negative class, indicating that this mitigation does not fully resolve class-dependent miscalibration. (Dark-bold: p<0.05 , Wilcoxon signed-rank test, 5 seeds; Light-bold: highest mean, not significant)
Clinical Condition
MedFuse (Loss-Upweighting)
DrFuse (Loss-Upweighting)
MeTra (Loss-Upweighting)
BS↓
BSc=1↓
BSc=0↓
BS↓
BSc=1↓
BSc=0↓
BS↓
BSc=1↓
BSc=0↓
1
20.22 ± 0.47
19.19 ± 1.02
20.71 ± 1.16
21.30 ± 0.55
20.82 ± 1.21
21.52 ± 1.34
21.35 ± 0.81
22.66 ± 3.41
20.73 ± 2.78
2
10.68 ± 0.51
17.41 ± 0.75
10.10 ± 0.61
10.46 ± 1.44
16.47 ± 2.92
10.00 ± 1.77
12.87 ± 1.95
11.96 ± 1.37
12.94 ± 2.23
3
20.79 ± 1.45
23.15 ± 2.02
20.55 ± 1.80
24.24 ± 4.12
21.09 ± 5.20
24.56 ± 5.03
20.63 ± 4.35
25.50 ± 8.98
20.14 ± 5.71
4
22.49 ± 0.73
21.98 ± 2.67
22.81 ± 2.78
24.15 ± 0.56
22.95 ± 1.58
24.88 ± 1.85
23.48 ± 0.51
25.52 ± 3.00
22.23 ± 2.64
5
20.05 ± 1.22
20.14 ± 2.46
20.02 ± 2.35
22.08 ± 1.49
20.60 ± 2.84
22.54 ± 2.81
21.79 ± 2.29
23.77 ± 6.22
21.17 ± 4.94
Appendix
Table C.14 : Brier Score for Multimodal Models under Loss-Upweighting. Aggregate and class-stratified Brier scores are reported for multimodal architectures trained with loss upweighting. Loss upweighting reduces positive-class Brier loss but increases negative-class and aggregate Brier loss, suggesting that the intervention shifts calibration error across classes rather than fully resolving the reliability gap. (Dark-bold: p<0.05 , Wilcoxon signed-rank test, 5 seeds; Light-bold: highest mean, not significant)
Figure C.1 : Unimodal CXR Stratified Calibration and Selective Prediction Across Clinical Conditions. For unimodal CXR, stratified positive-class ECE exhibits substantial variability across seeds and conditions (top), and AUROC/AUPRC remain below multimodal MedFuse performance. Across conditions, higher positive-class ECE AUC is associated with lower selective AUC (middle; Spearman’s rank p<0.05 ). Negative-class ECE AUC shows no clear monotonic relationship with selective AUC (bottom).
Figure C.2 : Unimodal EHR Stratified Calibration and Selective Prediction Across Clinical Conditions. Positive-class ECE is elevated for several conditions, indicating overconfidence on underrepresented positive cases. Compared to CXR, stratified ECE and AUROC/AUPRC show lower variability across seeds and conditions, yielding more stable predictions that closely mirror MedFuse behavior. Positive-class ECE AUC remains negatively associated with selective AUC, while negative-class ECE shows no clear monotonic relationship with selective performance.
Figure C.3 : MedFuse Stratified Calibration and Selective Prediction Across Clinical Conditions. Across conditions, higher positive-class ECE AUC consistently predicts lower selective AUC, indicating that overconfidence in positive predictions drives instability in selective evaluation. In contrast, negative-class ECE AUC remains low for most conditions and shows no clear monotonic relationship with selective AUC, suggesting that miscalibration is primarily driven by overconfident positive predictions rather than negative-class errors.
Figure C.4 : DrFuse Stratified Calibration and Selective Prediction Across Clinical Conditions. DrFuse exhibits elevated positive-class ECE across several conditions, indicating persistent overconfidence on underrepresented positive cases. Stratified ECE and selective AUROC/AUPRC closely mirror MedFuse trends, with higher positive-class ECE AUC associated with degraded selective performance. As in other models, negative-class ECE shows no clear monotonic relationship with selective AUC, suggesting that increased fusion complexity alone does not resolve calibration-driven reliability failures.
Figure C.5 : MeTRA Stratified Calibration and Selective Prediction Across Clinical Conditions. MeTRA shows class-dependent calibration behavior similar to MedFuse and DrFuse, with elevated positive-class ECE indicating overconfidence on underrepresented positives. Positive-class ECE AUC is negatively associated with selective AUC, while selective AUROC/AUPRC exhibit higher variance across seeds, consistent with other multimodal models, indicating that architectural complexity alone does not resolve calibration-driven selective instability.
Figure D1 : Comparative analysis for Acute Cerebrovascular Disease. For this condition, EHR predominantly drives the discriminative performance of MedFuse, while CXR shows poor positive-class calibration and contributes to the oscillatory behavior of selective AUROC. Applying label-weighted loss reduces early-threshold miscalibration in positive predictions but does not fully correct calibration at higher thresholds, leaving the overall selective prediction trend largely unchanged.
Figure D2 : Comparative analysis for Conduction Disorders. For this condition, the CXR model shows substantial seed-dependent variability, achieving higher mean performance but poorer calibration consistency than EHR. MedFuse integrates both modalities to improve selective AUROC and stabilize performance, though positive-class ECE remains variable. Label-weighted loss further reduces seed variance and attenuates positive-class ECE at higher confidence thresholds, yielding smoother and more reliable selective behavior, but with similar performance drops at the final thresholds.
Figure D3 : Comparative analysis for Respiratory Shock. For this condition, unimodal EHR and CXR achieve comparable baseline performance, with EHR showing slightly lower positive-class ECE at early thresholds. As rejection increases, however, EHR’s AUROC quickly deteriorates while CXR remains uniformly undercalibrated, and MedFuse largely inherits these calibration inconsistencies. Label-weighted loss substantially improves both stability and calibration, suppressing positive-class ECE spikes and producing a clearer monotonic gain in selective AUROC. This condition provides one of the strongest examples of how simple loss upweighting can directly reduce positive-class miscalibration and enhance the reliability of uncertainty estimates.
Figure D4 : Linear Regression of AUROC Against Stratified ECE. Relationship between AUROC and both positive- and negative-class ECE across 25 conditions. Regression slopes demonstrate a consistent, albeit not conclusive, negative association, indicating that poorer calibration modestly reduces discriminative performance in terms of AUROC.
Figure D5 : Linear Regression of AUPRC Against Stratified ECE. Relationship between AUPRC and class-stratified ECE across 25 conditions. The correlation between ECE and AUPRC exhibits a clear, monotonic decline, confirming via another approach, that high positive-class miscalibration strongly predicts lower precision-recall performance under selective prediction.
Safe predictions are a crucial requirement for integrating predictive models into clinical decision support systems. One approach to improving trustworthiness is to enable models to express uncertainty about individual predictions. However, current machine learning models frequently lack reliable uncertainty estimation, hindering real-world deployment. This limitation is particularly evident in multimodal settings, where models must effectively integrate heterogeneous information. In this work, we propose MedCertAIn, a predictive uncertainty framework that leverages multimodal clinical data to improve model performance and reliability for in-hospital mortality risk prediction as an indicator of patient deterioration. We design data-driven priors over neural network parameters using a hybrid strategy that considers cross-modal similarity in self-supervised latent representations and modality-specific data corruptions. We train and evaluate the models with such priors using clinical time-series and chest X-ray images from the publicly available datasets MIMIC-IV and MIMIC-CXR. Our results show that MedCertAIn achieves competitive predictive performance and substantial gains in selective prediction compared with the evaluated deterministic and stochastic baselines. These findings highlight the promise of data-driven priors in advancing robust, uncertainty-aware AI tools for high-stakes clinical applications. Our implementation is publicly available at: https://github.com/jlaitue/medcertain.
L. Julián Lechuga López, Tim G. J. Rudner, Farah E. Shamout
Tandon School of Engineering, New York University · New York University Abu Dhabi · University of Toronto
Selective prediction systems can mitigate harms resulting from language model hallucinations by abstaining from answering in high-risk cases. Uncertainty quantification techniques are often employed to identify such cases, but are rarely evaluated in the context of the wider selective prediction policy and its ability to operate at low target error rates. We identify a model-dependent failure mode of entropy-based uncertainty methods that leads to unreliable abstention behaviour, and address it by combining entropy scores with a correctness probe signal. We find that across three QA benchmarks (TriviaQA, BioASQ, MedicalQA) and four model families, the combined score generally improves both the risk--coverage trade-off and calibration performance relative to entropy-only baselines. Our results highlight the importance of deployment-facing evaluation of uncertainty methods, using metrics that directly reflect whether a system can be trusted to operate at a stated risk level.
Edward Phillips, Fredrik K. Gustafsson, Sean Wu +2
Department of Engineering Science, University of Oxford · Oxford Suzhou Centre for Advanced Research, University of Oxford, Suzhou
Multimodal prediction models add a data source, such as imaging or an assay, to clinical information, and health systems must decide whether that added test is worth acquiring. We audited 177 studies comparing multimodal with unimodal models: 81 from a systematic review and 96 from a probability sample of 2025-2026 PubMed records. Gains in the area under the receiver operating characteristic curve (AUROC) were smaller when a study fitted a separate model for every data source (adjusted difference, -0.030; 95% CI, -0.056 to -0.005). Within 78 such studies, leaving out the strongest source's model would have raised the average gain from 0.046 to 0.108. This complete benchmarking fell from 69.1% of review studies to an estimated 46.0% of recent ones. An estimated 22% assessed decision utility and none evaluated a model prospectively, yet most claimed potential clinical usefulness or improved care (84.0% and an estimated 87.7%).
Jiayu Feng
Department of Health Policy and Management, Harvard T.H. Chan School of Public Health, Boston, MA, USA