Shared and Private Factors of Death and Disability in Patient Representations

Public health measures the total burden of disease in disability-adjusted life years (DALYs), a single currency in which one unit is one year of full health lost. A DALY is the sum of two parts: years of life lost to early death, and years lived with disability, weighted by how severe that disability is. The sum assumes the two parts can be counted separately. Inside a patient, though, the same physiological decline both raises the risk of dying and erodes day-to-day function. Clinical deep learning models predict death and disability-like outcomes from one shared internal representation, so they never say how much of the signal behind each outcome is common and how much is specific to it. We propose to train patient models whose internal representation is explicitly split into three pieces. One is a shared "common decline" factor. A second is used only for predicting death, and a third only for predicting non-fatal morbidity. The three pieces are kept geometrically orthogonal, meaning statistically uncorrelated across patients, so that information cannot quietly migrate between them.

There are two payoffs. The first is a measurement: how large the shared piece must be gives an empirical estimate of how far the additive DALY bookkeeping misdescribes patient physiology. The second is a falsifiable engineering claim. When an intervention or a change of setting alters the coupling between death and disability, a model with separated factors should degrade less than an entangled one. The SPRINT trial is the motivating case: intensive blood-pressure control cut mortality but raised serious adverse events, so the two outcomes moved in opposite directions.

Core Claim

The obvious way to combine these directions is to split the latent space into two mutually orthogonal halves, one for death and one for disability, each read by its own head. The research file's own gap analysis proposes exactly this. We think that design is mis-specified. Death and morbidity are strongly correlated: in SPRINT, people who moved from robust to frail had roughly twice the all-cause mortality hazard (HR=2.03) Changes in frailty, intensive blood pressure treatment, and risks of adverse clinical outcomes. A strict two-way orthogonal split forces that common driver into one half, which starves the other head or makes it duplicate the factor.

The structure that matches the problem is the shared/private factorization of Factorized Orthogonal Latent Spaces. That paper separates the signal common to two camera views from view-specific signal using an inner-product penalty and a rank penalty that prunes unused dimensions. We reuse that machinery. The "views" become the two DALY components, and the inputs are patient time series rather than images. Our claim has two parts. The shared/private model will match an unconstrained multi-task model on standard outcome metrics. It will also beat it on individualized DALY estimates when the death–disability coupling shifts between training and testing, and it will beat the strict two-way split everywhere.

Research Program

The input is the hourly ICU time series produced by MIMIC-Extract, which reduces MIMIC-III to 104 clinical variables plus intervention indicators (ventilation, vasopressors) for adult first ICU stays of 12 hours to 10 days.

A recurrent or transformer encoder maps the first 24 hours of a stay to three codes:

  • a shared code $s$;
  • a mortality-private code $p_Y$;
  • a morbidity-private code $p_D$.

The mortality head sees only $(s, p_Y)$ and predicts death within a fixed horizon. The morbidity head sees only $(s, p_D)$ and predicts morbidity burden among survivors: ventilation hours, vasopressor hours and length of stay. Stacking a batch of codes into matrices $S$, $P_Y$, $P_D$ (one row per patient), we minimize

$$\mathcal{L} = \mathcal{L}_{Y} + \mathcal{L}_{D} + \lambda\left(\|S^\top P_Y\|_F^2 + \|S^\top P_D\|_F^2 + \|P_Y^\top P_D\|_F^2\right) + \gamma\, R(S, P_Y, P_D).$$

Here:

  • $\mathcal{L}_Y$ and $\mathcal{L}_D$ are the two outcome losses;
  • the bracketed term is the cross-covariance (inner-product) penalty from FOLS, where $\|\cdot\|_F$ is the Frobenius norm;
  • $\lambda$ sets how strongly the codes are pushed toward orthogonality;
  • $R$ is a log-singular-value rank penalty with weight $\gamma$, which lets the data choose how many dimensions each code keeps.

We run singular value decompositions only on small code matrices, so the cubic cost that FOLS warns about stays negligible.

A second variant enforces the split inside intermediate layers with the Cayley-parameterized orthogonal projection layer of Layer-Centric Factors of Variation Disentanglement. That paper reports that constraining only a bottleneck leaves intermediate layers tangled. Whether this holds for clinical time series is itself one of our questions.

To connect to DALY accounting, each patient gets an individualized health-loss target over a horizon after admission. The life-lost part is the probability of death multiplied by remaining reference life expectancy at the patient's age, using the standard-table approach described by Esser et al.. The disability part is morbidity duration multiplied by a disability weight. The research file supplies no disability weights for ICU states, so we treat the weight as a parameter and report every DALY-level result across a range of weights rather than claiming one value.

The scientific readout is the dimension the rank penalty leaves in $s$ relative to $p_Y$ and $p_D$, together with how much predictive accuracy each head loses when $s$ is ablated. A near-empty shared code would mean the additive DALY bookkeeping is benign at the patient level. A dominant one would quantify the overlap that the additive sum hides.

Baselines

The naive reference point is the formula itself with no learning: an age- and sex-only estimate. Expected life lost comes from reference life expectancy and a cohort-average mortality rate, and expected disability from cohort-average morbidity duration. Any learned model must beat this on per-patient DALY error, or the representation adds nothing.

The strongest existing method is an unconstrained multi-task recurrent model on MIMIC-Extract features, with one shared backbone and two heads. For mortality, MIMIC-Extract reports AUROCs between 0.85 and 0.90, and AUPRC often below 0.50 for rare events. Our model must stay within that AUROC band to show the constraint costs nothing on the standard task. For individualized DALY prediction there is no reported result found on Lacuna, so the multi-task model's own DALY error, run by us, is the bar.

The closest prior design is the biological-age model of Development of deep biological ages aware of morbidity and mortality. It trains a transformer autoencoder with a morbidity contrastive loss, a mortality loss, and a consistency loss that keeps its biological-age-gap features independent of chronological age. It treats morbidity and mortality as joint supervision on one representation; we instead separate them. We re-implement it on the same features and report it as a separate competitor. Its only reported comparison is a Kaplan–Meier separation at p<0.001, with no hazard ratios or AUCs given, so we have no number to compare against directly.

The ablations isolate the key ingredient:

  • the full model with $\lambda=0$, which keeps the three-code architecture but drops orthogonality;
  • the strict two-way orthogonal split with no shared code;
  • separate single-outcome models;
  • a β-VAE latent (a variational autoencoder with an extra weight on its prior penalty) with the same heads.

Every arm uses the same patient-level splits, the same encoder family and parameter budget within 5%, the same 24-hour input window, and the same number of hyperparameter trials (we plan 40, including tuning $\lambda$ and $\gamma$ for our arms). Results are reported over five seeds. Beating the multi-task baseline means a smaller per-patient DALY error under coupling shift, with seed-level confidence intervals that exclude zero, without falling below its in-distribution mortality AUROC.

Why not just…?

Why not just train an ordinary multi-task model with two heads? This is the main threat, and the in-distribution comparison may well be a tie. Shared backbones are often fine when training and test conditions match. Our claim rests entirely on the coupling-shift experiment. If the multi-task model degrades no more than ours when the death–disability correlation changes, the orthogonality constraint is not worth adopting, and we will report that. The evidence that it might matter is indirect: TS-Fingerprint Learning Fingerprints for Medical Time Series found that unconstrained masked autoencoders on ECG and EEG mixed physiology, patient identity and noise into one vector. That paper reports a +13.1% relative F1 gain on the ADFTD EEG dataset with pre-training versus training from scratch, and its ablations attribute the gain to the diversity term. It is a different task and does not establish what happens for survival outcomes.

Why not just train unconstrained and find shared and private directions afterwards with CCA or a projection? This is a serious alternative, and we include it as an arm. We run canonical correlation analysis (CCA) on the two heads' penultimate features from the multi-task model, then project out the shared directions. FOLS showed that linear CCA misassigns correlated noise to the shared space on synthetic multi-view signals, which is reason to expect post-hoc separation to be leakier. That evidence comes from a different domain, though. If the post-hoc arm matches our model under shift, the measurement part of the program still holds, and we would recommend the cheaper post-hoc procedure.

Why not just train two separate models, one per outcome? Separate models cannot leak information between outcomes, but neither can they reuse the common-decline signal. That should make them less sample-efficient. We test this by training every arm on 10%, 30% and 100% of the training cohort. Separate models winning at full data while losing at small data would itself be informative about how large the shared factor is.

Why not just use the strict orthogonal morbidity/mortality split the gap analysis suggests? This is the obvious idea, and our hypothesis is that it fails because the outcomes are correlated. It is an ablation, not a strawman. If it matches the shared/private model, then the shared factor is small and the additive DALY decomposition is nearly correct at the patient level, which would be a clean scientific result in itself.

Proposed Experiments

The decisive check is a coupling-shift evaluation, run in two forms.

The natural shift splits MIMIC-III by intensive care unit type. We train on medical-ICU stays and test on cardiac surgical stays, where mortality is generally low but post-operative support is common, and the reverse. Our hypothesis is that the death–morbidity relation differs between these units; we will measure it on the training data before running the split.

The controlled shift resamples the test set so that the correlation between death and morbidity burden is raised, removed, or reversed, while each outcome's marginal rate is held fixed. Prediction one: the multi-task and λ=0 arms degrade on per-patient DALY error in proportion to the size of the shift, and the orthogonal shared/private model degrades less. Prediction two: the strict two-way split is worse in distribution on both heads.

Alongside the shift tests we report:

  • per-head AUROC, AUPRC and calibration;
  • per-patient DALY mean absolute error across the range of disability weights;
  • a leakage probe, in which a linear classifier tries to predict death from $p_D$ alone and should perform at chance once $s$ is controlled for;
  • the shared-code dimension chosen by the rank penalty.

The bottleneck-penalty and layer-wise variants are compared under the same protocol.

If access to SPRINT trial data can be obtained, we will repeat the analysis there. The 12-month change in frailty index is the morbidity target and all-cause mortality the death target. The trial's treatment arm supplies a real intervention that moved the two outcomes in opposite directions: intensive control lowered mortality (HR=0.54 in stable-robust participants) and raised serious adverse events (HR=1.22). Under our hypothesis, a correctly separated model should place the treatment's predicted effect mainly in the private codes rather than in the shared factor.