- Rating
- 3
- Confidence
- 3
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
This paper studies the problem of unvailable modalities in the deployment of survival prediction methods. Other approaches address this problem by reconstruction or presentation matching. This may require learning variation beyond the survival task. The authors used a teacher trained on the auxiliary modality alone and take its survival outputs as the sole source of supervision from that modality and figured that existing survival distillation objectives typically encourage numerical agreement, although teachers and students observing different modalities need not share the same optimal predictions. They then prove that matching the teacher’s event probabilities at each time interval can shift the student’s optimal predictions. The authors then propose Ordinal Privileged Distillation (OPD), this method can transfer pair-wise rish ordering derived from the teacher's survival curves without matching its absolute score scale or pairwise score gaps. A pairwise logistic loss supplies ordinal supervision alongside a survival loss that accounts for right censoring, leaving the student’s inference architecture unchanged. Experimental results shows the method is effective.
Critical strengths
-
The idea is simple but effective. Th student learns pairwise risk ordering from single modal RNA teacher by pairwise loogistic loss. It does not introduce any incremental changes in network structure or training framework, instead, it is a purely loss-driven distillation paradigm that operates strictly at the output risk level.
-
The authors insightfully observe that 'the teacher and student condition on different inputs... so their optimal predictions can differ'. Proposition 2 proves that output matching requires an unmet compatibility condition which naturally motivates ordinal supervision that is 'invariant to strictly increasing transformations' without forcing the student to match 'numerical risk gaps'.
Critical weaknesses
- The pairwise loss scales quadratically within each batch, relying on an effective batch size. The paper currently lacks ablation experiments on sensitivity across varying effective batch sizes.
Questions for authors
-
The pairwise loss scales as and is tested with an effective batch size of . In memory-constrained scenarios where the batch size is much smaller (e.g., or ), does the reduction in pair diversity degrade the batch standardization and ordinal supervision? Have the authors evaluated the model's sensitivity to smaller batch sizes?
-
Since the RNA teacher is trained on right-censored data, how reliable are its continuous risk scores for pairs with very short follow-up times and early censoring? Would adding a confidence threshold or margin filter—to supervise only high-certainty pairs—improve stability?
3 comments
Full per-cohort results for the batch-size study. is the actual batch size; is the main setting.
| Model | BRCA | CRC | GBMLGG | HNSC | KIRC | LUAD | SKCM | UCEC | Mean | |
|---|---|---|---|---|---|---|---|---|---|---|
| ABMIL | 8 | 0.645±0.050 | 0.550±0.083 | 0.763±0.038 | 0.573±0.080 | 0.657±0.060 | 0.550±0.033 | 0.557±0.053 | 0.643±0.204 | 0.617 |
| ABMIL | 16 | 0.644±0.050 | 0.559±0.084 | 0.760±0.042 | 0.577±0.084 | 0.658±0.061 | 0.558±0.037 | 0.558±0.051 | 0.649±0.186 | 0.620 |
| ABMIL | 32 | 0.641±0.057 | 0.555±0.088 | 0.761±0.042 | 0.571±0.077 | 0.656±0.067 | 0.553±0.036 | 0.557±0.046 | 0.649±0.197 | 0.618 |
| ABMIL | 64 | 0.644±0.058 | 0.557±0.103 | 0.767±0.044 | 0.577±0.058 | 0.654±0.073 | 0.556±0.040 | 0.555±0.043 | 0.647±0.191 | 0.620 |
| ABMIL + OPD | 8 | 0.660±0.055 | 0.645±0.068 | 0.787±0.030 | 0.631±0.065 | 0.676±0.068 | 0.596±0.031 | 0.620±0.047 | 0.711±0.145 | 0.666 |
| ABMIL + OPD | 16 | 0.665±0.050 | 0.642±0.069 | 0.783±0.032 | 0.633±0.070 | 0.678±0.070 | 0.593±0.031 | 0.623±0.046 | 0.719±0.163 | 0.667 |
| ABMIL + OPD | 32 | 0.668±0.050 | 0.642±0.065 | 0.783±0.031 | 0.637±0.076 | 0.675±0.074 | 0.597±0.028 | 0.623±0.045 | 0.717±0.170 | 0.668 |
| ABMIL + OPD | 64 | 0.665±0.056 | 0.640±0.093 | 0.780±0.034 | 0.636±0.079 | 0.676±0.075 | 0.591±0.029 | 0.623±0.043 | 0.713±0.176 | 0.666 |
| TransMIL | 8 | 0.641±0.070 | 0.636±0.054 | 0.765±0.042 | 0.613±0.043 | 0.666±0.088 | 0.573±0.042 | 0.582±0.036 | 0.604±0.178 | 0.635 |
| TransMIL | 16 | 0.677±0.074 | 0.632±0.049 | 0.766±0.043 | 0.615±0.045 | 0.665±0.086 | 0.574±0.039 | 0.576±0.042 | 0.646±0.178 | 0.644 |
| TransMIL | 32 | 0.672±0.049 | 0.638±0.042 | 0.762±0.046 | 0.612±0.031 | 0.663±0.092 | 0.577±0.024 | 0.584±0.047 | 0.638±0.187 | 0.643 |
| TransMIL | 64 | 0.676±0.060 | 0.639±0.035 | 0.764±0.051 | 0.612±0.039 | 0.661±0.091 | 0.579±0.034 | 0.583±0.052 | 0.634±0.187 | 0.644 |
| TransMIL + OPD | 8 | 0.668±0.065 | 0.656±0.017 | 0.775±0.043 | 0.623±0.029 | 0.683±0.072 | 0.606±0.031 | 0.606±0.039 | 0.702±0.119 | 0.665 |
| TransMIL + OPD | 16 | 0.689±0.074 | 0.663±0.014 | 0.777±0.036 | 0.623±0.030 | 0.681±0.075 | 0.606±0.036 | 0.604±0.046 | 0.715±0.122 | 0.670 |
| TransMIL + OPD | 32 | 0.687±0.047 | 0.666±0.030 | 0.774±0.036 | 0.626±0.028 | 0.682±0.077 | 0.605±0.036 | 0.602±0.043 | 0.710±0.138 | 0.669 |
| TransMIL + OPD | 64 | 0.683±0.048 | 0.664±0.024 | 0.777±0.042 | 0.622±0.041 | 0.682±0.076 | 0.607±0.039 | 0.607±0.043 | 0.717±0.149 | 0.670 |
| MMP-Histo | 8 | 0.625±0.110 | 0.671±0.113 | 0.733±0.040 | 0.592±0.055 | 0.627±0.078 | 0.589±0.054 | 0.599±0.063 | 0.664±0.183 | 0.638 |
| MMP-Histo | 16 | 0.626±0.099 | 0.673±0.089 | 0.733±0.059 | 0.611±0.060 | 0.626±0.080 | 0.583±0.062 | 0.601±0.062 | 0.660±0.194 | 0.639 |
| MMP-Histo | 32 | 0.629±0.101 | 0.671±0.100 | 0.731±0.098 | 0.610±0.044 | 0.623±0.078 | 0.589±0.054 | 0.604±0.073 | 0.664±0.160 | 0.640 |
| MMP-Histo | 64 | 0.630±0.060 | 0.674±0.104 | 0.732±0.103 | 0.609±0.055 | 0.626±0.057 | 0.586±0.055 | 0.610±0.078 | 0.666±0.158 | 0.642 |
| MMP-Histo + OPD | 8 | 0.683±0.059 | 0.693±0.062 | 0.777±0.036 | 0.642±0.081 | 0.711±0.060 | 0.620±0.060 | 0.641±0.065 | 0.794±0.132 | 0.695 |
| MMP-Histo + OPD | 16 | 0.680±0.063 | 0.692±0.077 | 0.780±0.044 | 0.641±0.069 | 0.710±0.061 | 0.628±0.062 | 0.651±0.075 | 0.786±0.151 | 0.696 |
| MMP-Histo + OPD | 32 | 0.682±0.073 | 0.691±0.087 | 0.789±0.041 | 0.647±0.073 | 0.713±0.062 | 0.631±0.058 | 0.645±0.080 | 0.795±0.134 | 0.699 |
| MMP-Histo + OPD | 64 | 0.683±0.069 | 0.699±0.081 | 0.783±0.033 | 0.640±0.073 | 0.713±0.062 | 0.630±0.057 | 0.651±0.079 | 0.796±0.109 | 0.699 |
| PANTHER | 8 | 0.685±0.095 | 0.689±0.103 | 0.763±0.038 | 0.588±0.062 | 0.686±0.064 | 0.613±0.045 | 0.617±0.049 | 0.624±0.169 | 0.658 |
| PANTHER | 16 | 0.681±0.100 | 0.684±0.113 | 0.768±0.036 | 0.581±0.069 | 0.686±0.065 | 0.618±0.050 | 0.610±0.057 | 0.620±0.191 | 0.656 |
| PANTHER | 32 | 0.689±0.098 | 0.676±0.123 | 0.768±0.034 | 0.584±0.076 | 0.684±0.069 | 0.616±0.051 | 0.616±0.062 | 0.625±0.151 | 0.657 |
| PANTHER | 64 | 0.683±0.110 | 0.680±0.104 | 0.764±0.031 | 0.586±0.075 | 0.683±0.069 | 0.619±0.048 | 0.618±0.064 | 0.623±0.109 | 0.657 |
| PANTHER + OPD | 8 | 0.680±0.067 | 0.693±0.106 | 0.781±0.028 | 0.633±0.046 | 0.717±0.053 | 0.627±0.056 | 0.640±0.029 | 0.718±0.159 | 0.686 |
| PANTHER + OPD | 16 | 0.686±0.080 | 0.696±0.092 | 0.784±0.034 | 0.636±0.053 | 0.716±0.054 | 0.628±0.056 | 0.642±0.024 | 0.718±0.169 | 0.688 |
| PANTHER + OPD | 32 | 0.683±0.070 | 0.692±0.100 | 0.788±0.033 | 0.634±0.052 | 0.714±0.055 | 0.622±0.061 | 0.642±0.017 | 0.717±0.195 | 0.687 |
| PANTHER + OPD | 64 | 0.688±0.081 | 0.690±0.104 | 0.782±0.031 | 0.631±0.048 | 0.710±0.054 | 0.625±0.054 | 0.648±0.026 | 0.719±0.213 | 0.687 |
Q2. Since the RNA teacher is trained on right-censored data, how reliable are its continuous risk scores for pairs with very short follow-up times and early censoring? Would adding a confidence threshold or margin filter—to supervise only high-certainty pairs—improve stability?
A censoring-aware training likelihood does not guarantee that every teacher comparison is reliable. In particular, an early-censored outcome may not establish a patient's order relative to another patient, whereas the teacher can still predict an order from their RNA profiles. We tested the strictest high-certainty filter that the outcomes allow: keeping only Harrell-comparable pairs, in which the patient with the earlier observed time must have an observed event. This excludes censored–censored pairs and pairs whose earlier observation is censored, while keeping the RNA teacher as the source of the comparison labels:
| Model (pairs used) | BRCA | CRC | GBMLGG | HNSC | KIRC | LUAD | SKCM | UCEC | Mean | Mean std |
|---|---|---|---|---|---|---|---|---|---|---|
| ABMIL (Harrell pairs) | 0.665±0.056 | 0.646±0.082 | 0.776±0.042 | 0.644±0.074 | 0.675±0.078 | 0.591±0.031 | 0.621±0.045 | 0.672±0.192 | 0.661 | 0.075 |
| ABMIL (all pairs) | 0.665±0.056 | 0.640±0.093 | 0.780±0.034 | 0.636±0.079 | 0.676±0.075 | 0.591±0.029 | 0.623±0.043 | 0.713±0.176 | 0.666 | 0.073 |
| TransMIL (Harrell pairs) | 0.682±0.051 | 0.662±0.027 | 0.775±0.049 | 0.634±0.046 | 0.679±0.078 | 0.606±0.031 | 0.602±0.035 | 0.695±0.140 | 0.667 | 0.057 |
| TransMIL (all pairs) | 0.683±0.048 | 0.664±0.024 | 0.777±0.042 | 0.622±0.041 | 0.682±0.076 | 0.607±0.039 | 0.607±0.043 | 0.717±0.149 | 0.670 | 0.058 |
| MMP-Histo (Harrell pairs) | 0.682±0.071 | 0.685±0.096 | 0.783±0.060 | 0.639±0.074 | 0.700±0.073 | 0.633±0.050 | 0.647±0.078 | 0.706±0.162 | 0.684 | 0.083 |
| MMP-Histo (all pairs) | 0.683±0.069 | 0.699±0.081 | 0.783±0.033 | 0.640±0.073 | 0.713±0.062 | 0.630±0.057 | 0.651±0.079 | 0.796±0.109 | 0.699 | 0.070 |
| PANTHER (Harrell pairs) | 0.694±0.094 | 0.689±0.090 | 0.781±0.034 | 0.629±0.045 | 0.707±0.061 | 0.617±0.052 | 0.637±0.029 | 0.689±0.213 | 0.680 | 0.077 |
| PANTHER (all pairs) | 0.688±0.081 | 0.690±0.104 | 0.782±0.031 | 0.631±0.048 | 0.710±0.054 | 0.625±0.054 | 0.648±0.026 | 0.719±0.213 | 0.687 | 0.076 |
Restricting OPD to Harrell-comparable pairs does not improve accuracy or stability: performance changes little on seven cohorts and drops noticeably on UCEC (0.672 vs. 0.713 with ABMIL; 0.706 vs. 0.796 with MMP-Histo). The reason is fundamental to privileged distillation: even when a patient is right-censored early in follow-up, their transcriptomic profile is fully observed. Because the RNA teacher maps to a full survival curve and is trained across the cohort with a censoring-aware NLL, its continuous risk score provides an informative molecular risk ordering for censored pairs whose raw follow-up times alone are uninformative, as in UCEC ().
We hope this addresses your concerns, and we are glad to answer any further questions.
Thank you for the positive evaluation and the practical questions.
We retrained all four backbones, with and without OPD, at actual batch sizes of 8, 16, and 32, and compared them with the main batch-64 results:
For each backbone, the result with OPD varies by at most 0.5 points across batch sizes, and the gain over NLL-only training at the same batch size stays at about 4 points. The ordinal signal and the batch standardization remain effective.