Ad-Infer: Integrated Clinical Inference
SYSTEM-LEVEL ARCHITECTURE
Ad-Infer is an integrated clinical inferential system rather than a collection of independent diagnostic modules. Its EVAL-KBS architecture combines explicit domain knowledge, Type-2 causal reasoning, fuzzy evaluation, Hopfield-based recurrent neural computation, learning processes, deep-learning mechanisms and Bayesian belief updating.
| Domain knowledge Sleep-EVAL for sleep disorders; corresponding EVAL knowledge bases for other clinical domains. | Type-2 causal reasoning Clinical hypotheses are evaluated in relation to direct, indirect, contributory and competing explanations. | Fuzzy evaluation + learning Graded evidence, uncertainty, adaptive weights, neural learning and Bayesian belief updating. |
| Ad-Infer integrated inference Question selection, hypothesis generation, positive diagnosis, exclusions, differential diagnosis, temporal reasoning and longitudinal revision occur within the same inferential system. | ||
| Integrated language-model functions Ad-Infer uses language-model capabilities internally to extend semantic comprehension of the interview and of the relationship between new information and clinical causality. The interaction returns evidence continuously to the inferential process. | ||
A BIDIRECTIONAL CLINICAL LOOP
The language function does not sit outside Ad-Infer as a separate assistant. It is used by Ad-Infer after stimulation by the current clinical inference, causal reasoning and belief state. It supports adaptive questioning, reformulation, clarification and semantic interpretation.
The return path is clinically important. Ad-Infer can use information about questions that were posed or skipped, understood or misunderstood, accepted or refused, together with the content of answers and clarifications. Evidence that agrees with the current hypothesis can reinforce a positive diagnosis. Evidence that contradicts, qualifies or destabilizes the hypothesis can modify the belief model, trigger additional questions and become part of the differential-diagnosis process.
POSITIVE AND DIFFERENTIAL DIAGNOSIS
Ad-Infer does not stop when a recognizable diagnostic pattern appears. It evaluates whether the required elements are present, whether exclusion criteria are satisfied, whether another disorder or treatment better explains the findings, and whether comorbid diagnoses should also be considered. The system can therefore pursue several competing hypotheses and revise their relative support as information accumulates.
TEMPORAL AND LONGITUDINAL REASONING
Clinical meaning often depends on order and timing: which symptom appeared first, whether a treatment preceded a change, whether an event is persistent or episodic, and whether a later observation changes the interpretation of earlier information. Ad-Infer uses temporal information in the current interview and can revise prior inferences when new longitudinal data become available.
MATHEMATICAL AND FUZZY PROCESSING
Mathematical preprocessing supports transformations and comparisons required by the knowledge rules, including duration, age, timing and numerical ranges. Fuzzy evaluation allows symptoms, criteria and diagnostic support to be represented in graded form. Bayesian updating and learning processes alter the current belief state as new evidence is integrated.
SELECTED REFERENCES
- Ohayon MM. Sleep-EVAL, Knowledge Based System for the Diagnosis of Sleep and Mental Disorders. Canadian Intellectual Property Office, Registration #437699; 1994.
- Ohayon MM. Improving decision making processes with the fuzzy logic approach in the epidemiology of sleep disorders. J Psychosom Res. 1999;47:297-311.