Overview
Clinical Natural Language Processing (NLP) is an extraordinary technological achievement, capable of reading thousands of pages of medical text in a fraction of a second. However, it is fundamentally devoid of clinical judgment. Medical records are incredibly noisy. They contain speculative diagnoses, copy-pasted histories, hypothetical questions asked by family members, and symptoms explicitly ruled out by previous physicians. While our NLP engines are trained to handle negation and familial context, the sheer complexity of human language guarantees that an AI will occasionally misinterpret a chaotic clinical note. If these unverified AI extractions were allowed to feed directly into a diagnostic algorithm, the result would be catastrophic. A single false-positive symptom—for instance, the AI extracting "Seizures" from a note that actually reads, "The mother asked if the child's shivering could lead to seizures in the future"—would completely skew the Bayesian math. The algorithm would desperately try to find rare diseases that explain seizures, burying the true diagnosis deep beneath irrelevant noise. This is known as "diagnostic cascading," and it is the primary reason why fully automated, black-box diagnostic AI is clinically unviable. To prevent this, Sanjeevni enforces a strict, non-negotiable gateway. When the NLP engine extracts a symptom, it is placed in a "Pending Review" queue. Absolutely zero data enters the mathematical scoring matrix unless a verified, board-certified physician explicitly clicks the "Approve" button. This workflow ensures that the final scorecard is based exclusively on ground-truth, doctor-verified evidence, rather than algorithmic hallucinations.
The Necessity of Manual Review
Clinical Natural Language Processing (NLP) is an extraordinary technological achievement, capable of reading thousands of pages of medical text in a fraction of a second. However, it is fundamentally devoid of clinical judgment. Medical records are incredibly noisy. They contain speculative diagnoses, copy-pasted histories, hypothetical questions asked by family members, and symptoms explicitly ruled out by previous physicians. While our NLP engines are trained to handle negation and familial context, the sheer complexity of human language guarantees that an AI will occasionally misinterpret a chaotic clinical note.
The Approval and Rejection Mechanics
The user interface for the HPO review process is designed for maximum speed and absolute clarity. When a physician opens the Review HPO tab, they are presented with a list of all phenotypes extracted by the NLP. Next to each proposed Human Phenotype Ontology term is a snippet of the original source text, along with a hyperlink that jumps directly to the highlighted sentence in the uploaded PDF. This context is crucial; the physician must be able to see exactly why the AI thought the symptom was present before making a decision.
Manually Adding Missing Phenotypes
No matter how advanced the AI, it can only extract information that was actually written down. It is a common occurrence in rare disease medicine that a crucial, defining symptom was simply never documented in the patient's previous charts. A referring general practitioner might have missed a subtle facial dysmorphism, or failed to note a mild skeletal abnormality because they were focused entirely on an acute cardiac issue.