During 2003 an alpha draft of the program was ready and started using at our clinic. This has about 30 000 disease feature links for about 480 internal medicine diseases.
Thus it took over two decades for PAIRS to come to this stage. Enormous energy, time are spent to make this tool. Over a million lines of code is used during development of it. Many languages are learnt and used to develop and test it. It is hoped that it will be useful to doctors and other professionals to accurately diagnose their patients.
Medical diagnosis is a complex subject and suffers from several pitfalls. Although study of medicine is a science, practice is an art. Mistakes can happen at enormous cost to patient and their family and doctor. Clinical Decision Support Systems (CDSS) are developed to minimize the errors. AI-MED is designed to help doctors minimize errors in their practice. In a study it was found that 225 000 patients die each year because of medical error including diagnostic errors (15%) and side effects of medications (45%). CDSS is made mandatory for use in USA along with HIS to minimize these errors. Diagnostic errors are made by doctors because of several reasons. Psychologists studied these and found that salient distracting features might be one of the reasons. For example, one might think some features are important because of their current relationship to some event but may not be involved in disease process or unrelated to diagnosis. Similarly faulty reasoning might be due to cognitive or confirmation bias. Some other errors might be due to anchoring or framing or early closure of leads. AI-MED is designed to minimize these errors by disrupting the process. AI-MED diagnostic process is disruptive for traditional diagnosis (by not considering any bias invariably involved in human reasoning) and hence minimize errors.
Artificial intelligence (AI) consists of Natural Language Processing (NLP) and Diagnostic Decision Support (DDS) and are part of CDSS. Some examples of NLP include a statistical text classifier. However,clinical terms are much complex and usually are based on Latin and Greek terms. A Standardized Nomenclature of Medical Terminology-Clinical Terms (SNOMED-CT) was developed for text classification. The terms (over 300 000) are indexed by 9 digit numbers for accurate description and automated processing. Algorithms are built to use this index for correct interpretation of patient data. DDS is applied on patient data for diagnosis. Bayesian probabilistic belief networks are popular and their approximation methods can be used for diagnosis. Physician Assistant Artificial Intelligence Reference System (PAIRS) is developed in similar lines. It has about 28 000 disease-feature links for about 486 internal medicine diseases and 2000 features. PAIRS features consist of symptoms, signs or tests. It’s AI consists of NLP and DDS. NLP is based on SNOMED-CT word index analysis. It’s algorithm generates a word based indices from which the possible synonyms are selected and displayed. User can enter data as one likes and program looks for their synonyms from a feature list. AI-MED uses PAIRS database. User friendly NLP enables one to enter clinical data as one likes. For example, acronyms are identified correctly by NLP. Once patient data is entered, one can run DDS.
AI-MED uses Approximation method of Bayesian Probabilistic method for its DDS. This method was published in Journal of Artificial Intelligence Research by Tommi Jaakkola and Michael Jordan in 1999. Each of PAIRS features are weighted (0.09 to 0.99) according to their pathophysiological basis and their clinical importance. Diagnostic decision is clustered into one of each group for: infection, neoplasia, autoimmune or others. DDS runs on patient data to give a set of possible diagnoses. AI-MED gives diagnostic data irrespective of any bias. For any given patient data, it builds a case data from PAIRS database. Case data includes weights, incidence of disease and their statistical leak factors. DDS is designed to calculate an approximation of probability of a disease. This approximation has an upper and lower bounds. Accuracy of implementation of these algebraic algorithms are verified by resulting consistent numerical variation of 0.00004 to 0.00009 between the bounds. A Bayesian probabilistic estimation is made for a diagnosis. Finally, a set of investigations are suggested for testing the possible diagnosis. The output can be saved in a file for further reference.