Problem
A symptom checker must ask the patient clarifying questions the way a doctor does, then name a likely diagnosis. The system has to pick questions from about 5000 possible symptoms, while a patient has only one to five of them. Rule-based expert systems are accurate but every new symptom or diagnosis needs expert labeling. Classic models (naive Bayes, decision trees) are interpretable but not accurate enough. Reinforcement learning (RL) agents depend on the order of questions, struggle with a huge action space and are hard to train and deploy.
Example from the article: Sber built a symptom checker with its subsidiary DocDoc (Sber Health), now in a pilot.
Solution in brief
Two networks are trained jointly. The first predicts which symptoms to ask, the second predicts the diagnosis from the known symptoms plus the answers to the asked ones. A differentiable binarization passes the gradient from the diagnosis model back to the question model. Questioning stops when the entropy of the diagnosis distribution falls below a threshold.
Questions this case answers
- How to train an adaptive questioning system that picks which of thousands of sparse features to ask next and stops when its prediction is confident enough
- How to train a chain of dependent models jointly, keeping the intermediate outputs interpretable and preventing error accumulation across stages
- How to build a symptom checker that asks follow-up questions and stops once the diagnosis is confident?
- How to choose which of thousands of sparse symptoms to ask next without training a reinforcement learning agent?
Tools and methods
- Logical regularization — split a task into supervised subtasks and train the sub-models together with a weighted sum of losses.
- Gumbel-Softmax-style differentiable binarization — turns predicted symptom probabilities into a one-hot vector without blocking the gradient.
- Entropy of the diagnosis distribution — stopping rule for questioning.
- BERT / RoBERTa — production version: text tokens as input instead of ready symptom vectors.
Google or a one-time email code. No password.