Models that predict how long a cancer patient will survive now routinely combine a digitized pathology slide with a gene-expression profile. They are usually tested in one of two states: every input present and clean, or one input deleted outright. Real clinical data lives between those states. Sequencing is shallow, gene panels are truncated, biopsies are small, and blood draws are contaminated by cellular RNA. Optical wireless engineering has a precise language for this in-between regime. A laser link through air fades by degrees because of turbulence, fog, and beam misalignment. Engineers summarize a receiver's robustness by its diversity order, the slope at which its failure probability falls as signal quality improves. That slope tells you whether the link fails when any path fades or only when all paths fade at once.
This note proposes to import that measurement, and the reliability-weighted combining rule that achieves high diversity order, into multimodal survival analysis. The two-sentence version is this. We define a prognostic "outage" as a patient's predicted risk moving too far from what clean data would give, and we measure how fast outage probability falls as each input's quality improves. We then show whether a fusion rule that weights each input by a per-patient estimate of its reliability achieves the slope expected from truly independent inputs, where standard fusion fails as soon as the weakest input degrades.
Core Claim
Current multimodal survival models behave like receivers with diversity order one. Their predictions break when any single modality is moderately degraded, even though the other modality still carries enough prognostic signal. A fusion rule modeled on maximal-ratio combining should raise this slope toward the number of modalities. Maximal-ratio combining is the radio technique that weights each received copy of a signal by its estimated signal-to-noise ratio. In our version, the per-patient weight comes from a small quality estimator trained under graded, realistic degradations rather than all-or-nothing deletion. The optical side supplies the analysis tool and the combining principle. The oncology side supplies the application, the data, and the requirement that gains show up in calibration as well as ranking.
The formal object is simple. Let $q_m$ be the quality of modality $m$ for a given patient; examples are the fraction of sequencing reads retained or the fraction of tumor tissue tiles available. Let $\hat r(q)$ be the model's predicted risk at those qualities and $\hat r^{\ast}$ its prediction on clean inputs. We define
$$P_{\text{out}}(q) = \Pr\left(|\hat r(q) - \hat r^{\ast}| > \delta\right), \qquad d = -\lim_{\bar q \to \text{high}} \frac{\log P_{\text{out}}}{\log \bar q},$$
where $\delta$ is a clinically meaningful risk shift, for instance one that moves a patient across a risk-group boundary, and $\bar q$ scales all modality qualities together. The relation for $d$ is borrowed from the fading-channel result that selecting among $N$ independent Gamma-gamma branches yields order $N\min\{\alpha,\beta\}/2$, which is what Performance Analysis of Free-Space Optical Communication Systems With Multiuser Diversity Over Atmospheric Turbulence Channels derives for opportunistic scheduling. Diversity order grows with the number of independently fading paths. Our hypothesis is that concatenation-style fusion has $d$ near the single-modality value, while reliability-weighted fusion roughly adds the per-modality orders.
Why this, and what it is not
The file shows that no existing work connects free-space optics and cancer prognosis. The nearest work on each side is suggestive but incomplete.
Structural Prognostic Event Modeling for Multimodal Cancer Survival Analysis (SlotSPE) compresses slides and genomics into patient-specific "slots." It trains image slots to reconstruct genomics so the model survives complete genomic removal at test time. SlotSPE motivates this proposal, because it shows that missing modalities matter and can be handled. But it treats missingness as a binary mask and reports C-index only. We keep its architecture as a strong base and ask a different question: how does its output degrade under partial corruption, and at what slope?
Discordant Expression of Circulating microRNA from Cellular and Extracellular Sources measured whole-blood and plasma microRNA in 2,391 Framingham participants. The two compartments can be anticorrelated (miR-28-3p, $r=-0.24$). This constrains our degradation design. Real corruption includes structured contamination, not just additive noise or erasure, so our degradation suite must include a structured mixing corruption.
On the optical side, Ergodic Capacity Analysis of Free-Space Optical Links With Nonzero Boresight Pointing Errors separates a persistent offset from random jitter. That distinction maps onto systematic batch shift versus sampling noise in assays, and we test both.
The "fluctuation paradox" in the multiuser-diversity paper is the claim that more variance can help when you pick the best branch. We deliberately do not adopt it. Survival prediction averages over patients rather than selecting a best one. Selection-style fusion (use only the most reliable modality) therefore enters as a baseline, not as the proposal.
Research Program
We first build a graded degradation suite, one axis per modality, each tied to a real clinical failure mode.
For expression data, the suite has three corruptions. Binomial thinning of counts mimics shallow sequencing. Restriction to the genes on a targeted panel mimics truncated profiling. A mixing corruption blends each tumor's profile with an unrelated tissue-like background profile at a random proportion, which echoes the cellular masking seen in the Framingham analysis.
For slides, the suite has three corruptions. Random tile subsampling mimics small biopsies, blur and compression mimic scanning artifacts, and a fixed stain shift acts as the "boresight" offset. Each corruption has a scalar severity, and severities are drawn independently per patient from Gamma distributions. This is the standard turbulence parameterization, and it gives realistic heavy tails.
Second, we attach a quality estimator to each modality encoder of a SlotSPE-style model. This is a small head predicting how far that modality's embedding lies from its clean counterpart. It is trained with the same graded degradations, using the clean embedding as the target, which is available by construction.
At fusion time, each modality's contribution to the risk score is scaled in proportion to its estimated reliability, the inverse of its predicted error variance, as in maximal-ratio combining. The survival head is trained with the usual Cox partial likelihood from Deep Survival: A Deep Cox Proportional Hazards Network.
Third, we measure $P_{\text{out}}$ curves and the fitted slope $d$ for every model. We also report C-index, integrated Brier score, and calibration at fixed horizons. The last two follow the argument in Stop Chasing the C-index: This Is How We Should Evaluate Our Survival Models and Position: Stop Chasing the C-index when Evaluating Survival Analysis Models that ranking alone hides whether predicted risks are trustworthy. Outage is itself a calibration-flavored quantity: it asks whether the number handed to a clinician moves, not just whether patient order is kept.
Baselines
The naive reference is a linear Cox model on clinical covariates alone. It is unaffected by modality degradation and so sets the floor that any degraded multimodal model must stay above to be worth using. On METABRIC, a linear Cox model reached C-index 0.630 against DeepSurv's 0.643 in Deep Survival: A Deep Cox Proportional Hazards Network. This is the only absolute survival number in the file, and we will reproduce it before extending METABRIC with our expression degradations.
The strongest existing method is SlotSPE, run with its published architecture and its cross-modal reconstruction branch on the ten TCGA cohorts it used. SlotSPE reports outperforming prior methods in 8 of 10 cohorts with an overall 2.9% C-index improvement, and robustness under complete genomic removal. The file gives no absolute per-cohort C-index values, so there is no reported result on Lacuna for SlotSPE under graded degradation. We will rerun it ourselves on identical splits.
We also include LD-CVAE imputation, which SlotSPE beat under complete omission, to represent generic inpainting. A selection-diversity baseline uses only the modality with the higher estimated reliability for each patient. That baseline tests whether weighting beats choosing.
Every model gets the same backbones, the same patient splits, the same exposure to degraded training samples, and an equal hyperparameter search budget. Beating SlotSPE means three things. First, a fitted outage slope $d$ clearly higher than its value, with confidence intervals from patient bootstrap. Second, integrated Brier score no worse under moderate degradation. Third, clean-data C-index within noise of SlotSPE, so robustness is not bought with accuracy.
The key ablation keeps everything identical except the fusion weights, which are either fixed learned scalars or per-patient reliability estimates. A second ablation trains the full model with binary modality dropout only, instead of graded degradation.
Why not just…?
Why not just train with modality dropout? Randomly zeroing a modality during training is the obvious cheap fix, and it is a real threat, since it is roughly what gives SlotSPE its robustness to full removal. The weakness we expect is that dropout teaches the model to handle "absent," not "present but wrong." A moderately corrupted expression profile looks like a valid input, so the model has no signal to down-weight it, and the corrupted input may hurt more than an absent one would. The dropout-only ablation settles this directly. If dropout matches our slope and calibration under graded corruption, the reliability head is unnecessary, and we would report that the diversity-order measurement is the contribution, not the combiner.
Why not just use predictive uncertainty from ensembles or Monte Carlo dropout as the weight? Ensemble disagreement is a generic reliability signal and needs no degradation-specific training. It measures model uncertainty about the output, however, not how corrupted a specific input is. Ensembles can also agree confidently on systematically shifted inputs such as the stain-offset corruption. We will run ensemble-variance weighting as an alternative weight source inside the same combiner. If it matches our estimator on the offset corruptions as well as the noise corruptions, it wins on simplicity.
Why not just impute the degraded modality back to clean with SlotSPE's reconstruction branch? Reconstruction from the image is exactly what SlotSPE does, and it may already provide the diversity we are after. But it needs to know when to trust the reconstruction over the measured profile, and under partial corruption nothing tells it. The comparison of SlotSPE with and without our reliability weights, at matched degradation, shows whether reconstruction plus a crude missingness flag is enough.
Proposed Experiments and Decisive Checks
The main experiment trains all models on the ten TCGA cohorts with disease-specific survival labels. We evaluate each on a grid of severities crossing both modalities and plot $P_{\text{out}}$ against mean quality on log axes. We fit slopes in the high-quality region and report C-index and integrated Brier score at each grid point. METABRIC, which pairs clinical data with expression, serves as a second, larger two-modality test.
The decisive check has three parts. Our model's slope should approach the sum of the single-modality slopes, while concatenation fusion and SlotSPE should sit near the smaller single-modality slope. The gap should persist in integrated Brier score, not just in C-index. And it should appear for the structured mixing and offset corruptions, not only for additive noise.
A negative result is informative too. If every fusion method shows order near one even with oracle reliability weights, the two modalities are not carrying independent prognostic information. That would argue against adding costly modalities for robustness at all.
Finally, the degradation suite is synthetic, so we check real quality variation. We stratify held-out TCGA patients by naturally occurring quality proxies, such as available tumor tissue area and profile completeness. We then test whether estimated reliability tracks these proxies and whether the reliability-weighted model's advantage concentrates in the low-quality stratum. If the estimator only recognizes our synthetic corruptions, that will show here, and it is the result that decides whether the method is ready for clinical data.
References
- Performance Analysis of Free-Space Optical Communication Systems With Multiuser Diversity Over Atmospheric Turbulence Channels
- Structural Prognostic Event Modeling for Multimodal Cancer Survival Analysis
- Discordant Expression of Circulating microRNA from Cellular and Extracellular Sources
- Ergodic Capacity Analysis of Free-Space Optical Links With Nonzero Boresight Pointing Errors
- Deep Survival: A Deep Cox Proportional Hazards Network
- Stop Chasing the C-index: This Is How We Should Evaluate Our Survival Models
- Position: Stop Chasing the C-index when Evaluating Survival Analysis Models