DDXPlus

Distributional diagnosis over a closed set of 49 pathologies. The model reads a patient case — demographics, then a transcript of yes/no and categorical symptom and antecedent questions — and must output a probability distribution over the 49 conditions. Unlike the other domains, the gold target is a genuine distribution: DDXPlus ships a per-case differential with calibrated probabilities, not a single right answer.

Dataset entry

Demographics:
Age: 80
Sex: M

Symptoms:
Have you been coughing up blood? Yes
Do you have pain somewhere, related to your reason for consulting? Yes
Characterize your pain: a knife stroke
Do you feel pain somewhere? posterior chest wall(L)
How intense is the pain? 1
Does the pain radiate to another location? nowhere
How precisely is the pain located? 6
How fast did the pain appear? 4
Are you experiencing shortness of breath or difficulty breathing in a significant way? Yes
Do you have a cough that produces colored or more abundant sputum than usual? Yes
Do you have a fever (either felt or measured)? Yes
Have you had chills or shivers? Yes
Do you have skin lesions or redness related to your condition? Yes
What color is the rash? pink
Do your lesions peel off? Y
Is the rash swollen? 1
Where is the affected region located? back of the neck, thoracic spine, buttock(R), buttock(L), flank(R)
How intense is the pain caused by the rash? 4
Is the lesion larger than 1cm? N
How severe is the itching? 1
Have you noticed any new fatigue or general malaise? Yes
Do you have a cough? Yes

Antecedents:
Do you drink alcohol excessively? Yes
Do you have heart failure? Yes
Have you ever had pneumonia? Yes
Do you have COPD? Yes
Do you have asthma / used a bronchodilator? Yes
Have you had surgery within the last month? Yes
Have you traveled out of the country in the last 4 weeks? N
Is your BMI less than 18.5? Yes

Representations

The same 49-condition target, three output formats the model writes (real rollouts).

Verbalized hypothesis list. Named conditions with weights and a residual bucket.

<hypotheses>
<h1>Acute dystonic reactions</h1><w1>0.70</w1>
<h2>Myasthenia gravis</h2><w2>0.15</w2>
<h3>Guillain-Barré syndrome</h3><w3>0.10</w3>
<w_other>0.05</w_other>
</hypotheses>

Class-hierarchy DSL. Latent classes with priors, per-class disease weights (d) and symptom likelihoods (s+); the posterior over the 49 is computed in closed form.

<spec>
u 0.05
c cardiac_emergency=0.6
  d D36=0.45
  d D37=0.3
  d D47=0.2
  d D10=0.15
  s+ chest_pain=0.8
  s+ palpitations=0.9
  s+ shortness_of_breath=0.7
  s+ fear_of_dying=0.85
c psychiatric_emergency=0.4
  d D33=0.4
  d D08=0.3
  d D23=0.2
  s+ anxiety=0.8
  s+ numbness=0.7
  s+ detachment=0.7
</spec>

Probabilistic program. The model writes the evidence function; the harness runs inference for the posterior over diseases.

def evidence(dx):
    # 54M: hemoptysis, dyspnea, productive cough, COPD, cystic fibrosis, prior pneumonia
    pyro.sample("respiratory_symptoms", dist.Bernoulli(
        per_disease({0:.9, 11:.85, 13:.8, 34:.75, 44:.7, 37:.65}, default=.45)[dx]), obs=ONE)
    pyro.sample("chronic_lung_history", dist.Bernoulli(
        per_disease({0:.7, 11:.65, 13:.6, 34:.55, 37:.5}, default=.35)[dx]), obs=ONE)