A diagnostic model rules a disease out partly from what the patient does not have. A child without the characteristic whooping cough is less likely to have pertussis, so "no whoop" is a strong rule-out. When childhood vaccination pushes pertussis into adults, who often have only a persistent cough, that same absent symptom carries almost no information. A model trained mostly on pediatric cases will keep using it as a rule-out and will systematically miss the disease in the population that now carries most of the burden.
We propose to treat an absent finding the way the better recommender systems treat an unclicked item. Silence counts as evidence only to the degree that the signal would have appeared had the hypothesis been true. Concretely, we build a diagnostic model in which a disease is penalized for each absent symptom in proportion to how likely that symptom is to show up for that disease in a patient of this age and sex. That likelihood is learned by a small model shared across diseases, so a pattern such as "adults present with blunted symptoms" can carry over to diseases whose adult cases are rare in training. The claim is that this specific change prevents a measurable class of errors that age-aware models, prior correction, and negation filtering do not fix: wrongly ruling out diseases that have shifted into new age groups.
Core Claim
Demographic disease shift changes two things. The first is how common a disease is in each age group, which epidemiologists track and which standard label-shift correction can absorb. The second is what a disease looks like in each age group, and in particular which symptoms reliably appear. The second change mostly damages the negative half of diagnostic reasoning. Associative diagnostic models learn absent-symptom penalties from whatever age mix they were trained on, and they apply those penalties unchanged to new cohorts. Our hypothesis is that if each negative finding is weighted by a demographic-conditioned "would it have shown?" probability, kept in its own channel, the false rule-out rate on age-shifted diseases will fall, and in-distribution accuracy will not.
The mechanism comes from Signals in the Silence. That paper weights the zero entries in an interaction matrix (tasks a crowd worker did not take) by how available those tasks were, on the logic that ignoring something you never saw is not a rejection. Our translation is direct. An absent symptom is a "zero entry," and its weight should be the probability that the symptom would have been expressed, given the disease and the patient's demographics. With $a$ the patient's demographic group, $P$ and $N$ the sets of present and explicitly absent findings, $\pi_a(d)$ the prior for disease $d$ in group $a$, and $e_\phi(j,d,a)$ a learned probability that finding $j$ manifests when $d$ is present in group $a$, the disease score is
$$s(d \mid x) = \log \pi_a(d) + f_\theta(P, d, a) + \sum_{j \in N} \log\big(1 - e_\phi(j,d,a)\big).$$
Here $f_\theta$ is an unrestricted neural scorer for positive evidence. The last term is the negative channel. An absent finding costs little whenever $e_\phi$ says it would often be missing anyway. We factorize the expression model as $e_\phi(j,d,a) = \sigma(u_d^\top M_a v_j)$, where $u_d$ embeds the disease, $v_j$ embeds the finding, $M_a$ is a low-rank matrix for each demographic group, and $\sigma$ is the logistic function. Because $M_a$ is shared across all diseases, attenuation patterns learned from diseases with plenty of adult data carry over to diseases without it. Keeping the prior term separate means that base-rate shifts, which can be estimated from surveillance sources such as the Global Burden of Disease analyses, can be swapped in without retraining the evidence model.
Relation to Prior Work
Three papers motivate the gap. DDXPlus built a 1.3-million-patient synthetic differential-diagnosis benchmark. Its authors concede that their generator does not guarantee capture of "relevant negative evidence (symptoms the patient doesn't have), which is a key part of 'ruling out' diseases." DDxT feeds age and sex into a transformer that reaches near-ceiling accuracy on DDXPlus. However, it has no explicit representation of negated findings and assumes training and test share an age-disease distribution. We keep DDxT as the backbone for positive evidence and change only how absence is scored and how the test distribution is constructed.
Retrieval Shift as a Source of Demographic Bias in Medical RAG studies retrieval-augmented generation, in which a model answers using documents fetched by a retriever. It shows that adding race or gender tokens flips 18.87–34.92% of answers in such systems. Its prescription is demographic-invariant representations. That prescription suits attributes that are clinically irrelevant by construction, which is how their MedQA slice was filtered. It is the wrong prescription for age under disease shift, where demographics legitimately change both base rates and presentation. Our proposal takes the opposite position for this case: demographics should modulate evidence weights, through an explicit and inspectable term.
Two papers supply technique. Implicit Negative Feedback in Clinical Information Retrieval showed that treating negated clinical phrases as negative relevance feedback, by subtracting a score for the negated part of a lexical query, beats simply deleting them. This is our evidence that negations carry signal worth keeping. That work operates on keyword retrieval and ignores demographics. Recommendations with Negative Feedback via Pairwise Deep Reinforcement Learning keeps separate recurrent states for clicked and skipped items so that abundant negatives do not swamp sparse positives. We borrow that separation, but our negative channel is gated by exposure rather than learned freely.
One paper constrains the design. Cross-domain Transfer of Valence Preferences reports "negative transfer" when avoidance patterns are moved across mismatched domains. This is exactly the risk of sharing $M_a$ across diseases. We test it explicitly rather than assume transfer helps.
Research Program
The work has three parts. The first is a controlled shift benchmark built from DDXPlus. We select a set of target pathologies and resample training patients so that those diseases occur almost entirely in younger age bands. For held-out test patients in older bands, we mask designated "classic" findings as explicitly absent with probability $q$, imitating attenuated adult presentation; $q$ is swept from 0 to 0.8 as a dose variable. A second, "transfer" variant gives the training set adult cases with attenuated presentation for a disjoint set of donor diseases, which tests whether the shared expression model carries the pattern to target diseases it never saw in adults.
The second part is the model itself: DDxT-style positive scoring, the factorized negative channel, and a pluggable age-specific prior. It is trained with the standard differential-diagnosis objective. The DDXPlus authors found that predicting the full differential acts as a strong regularizer, so we keep that objective.
The third part is a real-data check on MIMIC-IV hospital records, processed with the open MIMIC-IV pipeline. We extract negated findings from discharge notes ("afebrile," "denies cough") with NegEx and link them to diagnosis codes for infections with known age-dependent presentation, such as endocarditis, tuberculosis and pneumonia. The first question is descriptive and cheap: does the likelihood ratio of a given negated finding for a given diagnosis actually vary across age bands? If it barely varies, the mechanism has little real-world purchase, and we will say so.
Baselines
The naive reference is DDxT trained and evaluated as published, with age and sex as encoder inputs and absent findings either omitted or encoded as ordinary tokens. On standard DDXPlus it reports GTPA@1 (the true pathology ranked first) of 99.98% and differential F1 of 0.9472 (DDxT). AARLC reaches differential recall of 97.73% with differential training (DDXPlus). We found no reported result on Lacuna for any model on an age-shifted DDXPlus split, so all shifted-split numbers will be produced here. DDxT is the strongest existing competitor because it already conditions on demographics. If age conditioning alone sufficed, it would show it.
The second baseline is DDxT plus label-shift correction, in which the output is reweighted by test-cohort age-specific priors that are given to it exactly, as an oracle. This isolates whether the problem is base rates. The third is DDxT with all negated findings filtered out, which mirrors the NegEx-filtering arm of the clinical retrieval study. The fourth is DDxT fine-tuned on $k$ labeled target-cohort patients, for $k$ in {10, 100, 1000}, which is the obvious practical fix.
All baselines share our training data, backbone size, hyperparameter search budget, and compute. The few-shot fine-tuning arm receives the same $k$ target examples our model may use to adjust $M_a$.
Two ablations isolate the key ingredient. The first replaces $e_\phi$ with a demographic-independent $e_\phi(j,d)$, to test whether demographic gating matters. The second replaces the factorized $M_a$ with a free per-(disease, age) table, to test whether sharing across diseases is what drives transfer.
Beating the baselines means two things at once. On the shifted split, our model must lower the false rule-out rate on target diseases in older patients relative to both DDxT with oracle priors and DDxT fine-tuned with the same $k$. It must also stay within 0.5 points of DDxT's GTPA@1 on the unshifted test set.
Why not just…?
Why not just give the model age as an input? DDxT already does, and it can learn age-specific rule-out behavior only for disease-age pairs it has seen. Under shift, the adult presentations of target diseases are by construction rare or absent in training. The demographic-independent-$e_\phi$ ablation and the plain-DDxT baseline on the transfer variant test this directly. If plain DDxT matches us there, the core claim fails.
Why not just correct the priors with surveillance data? Prior correction moves probability mass toward diseases that have become more common in a cohort. It cannot undo a learned penalty saying "this absent symptom makes the disease unlikely," because that penalty sits in the likelihood, not the prior. The oracle-prior baseline isolates exactly this. Our prediction is that it recovers part of the gap at $q = 0$ and progressively less as $q$ grows.
Why not just strip out negated findings? That avoids false rule-outs, but it throws away discrimination everywhere else. The clinical retrieval study found that subtracting negated content beats deleting it (Implicit Negative Feedback in Clinical Information Retrieval). We expect the filtered baseline to lose in-distribution differential F1, and its result on the shifted split shows how much of our gain could be bought simply by ignoring negatives.
Why not just fine-tune on a few adult cases? This is the real threat. With enough target data, fine-tuning should close the gap. Our claim is limited to sample efficiency: at small $k$, and in the transfer variant where no target-disease adult cases exist, a shared expression model should do better. The $k$-sweep settles where, if anywhere, that advantage lives.
Proposed Experiments and Decisive Checks
The primary metric is the false rule-out rate: the fraction of older-cohort target-disease patients whose true disease falls outside the top-$k$ differential. We plot it against the attenuation dose $q$. Secondary metrics are GTPA@1 and differential F1 on both shifted and unshifted splits, plus calibration (expected calibration error) of the target-disease probability by age band.
Three results would be decisive. The first is that the gap between our model and DDxT with oracle priors widens as $q$ increases, which is the signature that the improvement comes from correcting negative evidence rather than base rates. The second is that in the transfer variant, the factorized model beats the per-pair table on target diseases without hurting donor diseases. If it hurts donor diseases, that is the negative transfer CVPM warned about, and it would argue for a gated or disease-clustered $M_a$.
The third result comes from MIMIC-IV. The age-band likelihood ratios of negated findings, estimated from records, must vary substantially. If they do, a model trained on earlier admission years and tested on later ones, stratified by age, should reproduce the synthetic false rule-out pattern at a smaller scale. If they do not, the proposal reduces to a property of synthetic data, and the note's conclusion will report that. As a reference for real-world difficulty, the best full MIMIC-IV-ICD10 coding model reaches Macro F1 of only 5.42% (Mimic-IV-ICD). For that reason we restrict the MIMIC-IV analysis to a small set of frequent infection codes rather than the full label space.
References
- Signals in the Silence: Models of Implicit Feedback in a Recommendation System for Crowdsourcing
- DDXPlus: A New Dataset For Automatic Medical Diagnosis
- DDxT: Deep Generative Transformer Models for Differential Diagnosis
- Retrieval Shift as a Source of Demographic Bias in Medical RAG
- Implicit Negative Feedback in Clinical Information Retrieval
- Recommendations with Negative Feedback via Pairwise Deep Reinforcement Learning
- Cross-domain Transfer of Valence Preferences via a Meta-optimization Approach
- An Extensive Data Processing Pipeline for MIMIC-IV
- Mimic-IV-ICD: A new benchmark for eXtreme MultiLabel Classification