Development of Artificial Intelligence System

Development of Artificial Intelligence System for Differential Diagnosis of Diseases

1. Mark D.Kats ( born in 1937, higher education) is an applied mathematician, doctor of science, academician of Ukrainian Production Engineering Academy and Ukrainian Academy of Ecological Sciences, corresponding member of International Academy of Computer Sciences and Systems. Donetskaya St. 37/24 Severodonetsk Lugansk region 349940 Ukraine Tel.: (380-6452) -3-08-75, E-mail: (Kats M.D.).

2. Name of the project is

"DEVELOPMENT OF ARTIFICIAL INTELLIGENCE SYSTEM FOR COMPUTER-AIDED DIFFERENTIAL DIAGNOSIS OF DISEASES A...N DIFFICULT TO DIFFERENTIATE".

3. Application field of the offer Medicine.

Upon developing the project, the problem of correct and express differential diagnosis inside the group of diseases A...N difficult to differentiate using a formalized procedure realized on a computer will be solved.

4. Description of the offer.

A differential diagnosis model will be constructed in collaboration with a medical institution specialized in treatment of a certain class of diseases difficult to differentiate using the submitted experimental data, requirements to which are given below. After corresponding clinical verification and correction of the disease model a software package " ARTIFICIAL INTELLIGENCE SYSTEM FOR COMPUTER-AIDED DIFFERENTIAL DIAGNOSIS OF DISEASES A....N'' will be developed. After patenting it will be a final software product suitable for marketing in all countries. THE PROPOSED CONSTRUCTION METHOD FOR MODELS OF DIFFERENTIAL DIAGNOSIS OF DISEASES DIFFICULT TO DIFFERENTIATE IS A RESULT OF STRATEGIC INVESTIGATION WHICH AIM WAS TO DEVELOP A NEW MATHEMATICAL SIMULATION METHOD ENABLING TO CONSTRUCT INFORMATIVE MATHEMATICAL MODELS OF COMPLEX OBJECTS BELONGING TO CLASS OF "LARGE" SYSTEMS BASED ON EXPERIMENTAL DATA OBTAINED IN THE COURSE OF OBSERVATION ("PASSIVE" EXPERIMENT), USING A FORMALIZED PROCEDURE (WITHOUT AN EXPERT).

Mathematical simulation method (mosaic portrait method) based on artificial intelligence philosophy enabling to construct mathematical models of the object (system) under study, carrying an essential scope of new, untrivial, yet unknown information about relations between its input and output variables using table of experimental data which describe behavior of an object (system) of any physical nature, without additional a priori information about the model structure has been developed, evaluated and experimentally verified over many years. As applied to medical diagnosis, the mosaic portrait method allows to obtain for each of diseases being differentiated a set of differential syndromes, inherent only to a specific disease, the overwhelming majority of these syndromes were unknown before.

Differential diagnosis model construction problem can be stated as the following mathematical formulation.

GIVEN:

Experimental data M=XxY (X={Xij}, i=1,m, j=1,n; Y={Yil}, l=1,k) table, each line of which contains information about symptoms values (Xij) and verified diagnosis Yil by an i-th patient. (Here m is a number of lines (patients) in table M, n is a number of columns (symptoms) in table M, K is a number of diseases to be differentiated.)

IT IS NECESSARY: to construct a mathematical model allowing to reliably differentiate each of K of diseases to be differentiated, on the basis of the table M by means of formalized procedures.

If a table, each line of which contains symptoms values and verified diagnosis for one patient, has been obtained from the investigation results, and all observed patients belong to a set of diseases difficult to differentiate from each other, being studied, a mathematical model of differential diagnosis can be constructed using the mosaic portrait method by realization of the following formalized procedures:

- transformation of initial experimental data table M to table M' by transfer from continual scales used for measurement of symptoms to discrete ones. In this case a specific code is given to each values subrange of each of symptoms;

- search of codes combinations which can be found in table M' in lines belonging to one disease and not found in any line with other diseases.

These combinations are interpreted as differential syndromes of a corresponding disease. Thus, the mosaic model consists of K disjunctions, each containing differential syndromes of one of K of diseases to be differentiated. Number of differential syndromes is very large and the most of them are new, untrivial, unknown before in the medical science.

5. Advantages in comparison with analogues.

As applied to the problem of differential diagnosis of K of diseases, a mathematical model obtained using the mosaic portrait method contains k of subsets of differential syndromes, the most of them were unknown before. Mathematical methods which could be used for construction of differential syndromes of diseases to be differentiated during acceptable time period even at 15 symptoms, are unknown at present. For the mosaic model, there are practically no limitations as to number of symptoms.

6. Possible users.

Potential users of diagnosis software packages can be all medical institutions engaged in treatment.

7. Finality degree.

There is a software for mosaic portrait method realization which can be used to construct differential diagnosis model for any diseases difficult to differentiate if corresponding experimental data are available. There is a structure of the package to which a model of any differential diagnosis can be inserted. There are already operating diagnosis software packages (differential diagnosis: Gastric ulcer - Gastric cancer, Idiopathic hypertension - Nephritic hypertension, etc.), for which mathematical models have been developed in collaboration with Military Medical Academy (Saint-Petersburg). One of these packages realizes an intelligent system "Prognosis of myocardial infarction complications at the acute phase".

The following problems are solved by this package:

- prognosis of one or several possible myocardial infarction complications according to the information about patient state collected at the first day of his stay in a hospital;

- in case of prognosis of several complications , the probability of realization of each of them;

- recommendations on preventive therapy of a complication (complications) being prognosticated and corresponding symptoms-eliminating therapy taking into account compatibility of medicines and treatments;

- every two days - a new prognosis corrected on the basis of treatment results and recommendations on preventive and symptoms-eliminating therapy corresponding to this prognosis;

- on inquiry, displaying of information concerning differential syndromes on which basis the prognosis was done;

- input of information about the patient, complications prognosis, syndromes, on which basis the prognosis was done, and recommended treatment into the data base. Initial information about the patient state including anamnesis, examination, electrocardiograph data (total 39 symptoms) is entered by keystroke in question-answer mode. Experimental verification of efficiency of use of the obtained model in a cardiology clinic of Military Medical Academy (Saint-Petersburg), 23rd teaching hospital (Moscow), 20th city hospital (Saint-Petersburg), 42nd city hospital (Saint-Petersburg) has shown that accuracy of prognosis amounts to 85-88% and lethality of myocardial infarction can be reduced by purposeful preventive therapy by 36-45%.

8. Proposal protection.

The mosaic portrait method is not protected by patent but the main part of algorithm ensuring model completeness and polynomiality of computer time and realization time versus dimensionality of the problem, correspondingly, has not been published, and only the author is possessed of the software for its realization. The differential diagnosis and myocardial infarction complications prognosis models included into corresponding packages are not protected by patent.

9. Mark D. Kats has the property right to this offer.

Mark D.Kats, doctor of science, Donetskaya St. 37/24, Severodonetsk, Lugansk region, 349940 Ukraine, tel.: (380-6452)-3-08-75, E-mail: root@ixt.lg.ua (Kats M.D.).

Attachment to the project:

"DEVELOPMENT OF ARTIFICIAL INTELLIGENCE SYSTEM FOR COMPUTER-AIDED DIFFERENTIAL DIAGNOSIS OF DISEASES A...N DIFFICULT TO DIFFERENTIATE".

We propose MEDICAL INSTITUTIONS to take part in development of artificial intelligent systems which have no analogues worldwide enabling to carry out reliable differential diagnosis of each disease using mathematic models describing a great number of diseases under study which symptoms are similar.

PROGRAM OF WORKS ON DEVELOPMENT OF ARTIFICIAL INTELLIGENCE SYSTEM FOR COMPUTER-AIDED DIFFERENTIAL DIAGNOSIS WITHIN GROUP OF DISEASES A...N DIFFICULT TO DIFFERENTIATE.

1. Analysis of the problem status

Diagnosis is and will be the most important problem of medicine, and the accuracy of diagnosis achieved in certain historical periods determines mainly the state-of-the-art in medical science.

While examining a patient in modern diagnosis centers (anamnesis, physical examination, laboratory and instrument methods data, clinical data, etc.) a very large scope of initial data (more than 300 characteristics being measured mainly using numeral scales) is collected. If each of these characteristics is measured only using the most simple name-scales ("yes-no", "more-less") the quantity of initial data will make 2 bits to 300th power which is substantially higher than the number of elemental particles in all visible Universe.

As the human body is very complicated and it is characterized by practically infinite number of disease symptoms, symptoms and clinic of a disease are greatly influenced by the individual features of a patient and knowledge of specialists is limited, medical diagnosis, nowadays, is not a science but rather an art of a few highly qualified professionals.

After computers appeared and applied mathematics is developed, works related to attempts to formalize the diagnosis process using mathematical models boomed. The results of these works mainly did not come up to expectations and rare successes are connected either with relative simplicity of the problem (to differentiate diseases sufficiently remote from each other in the symptoms space) or with its inadequate simplification. As a result, at best models diagnosing a disease not worse than an average doctor appeared.

Principal difficulties in simulation of "large" systems (to which medical diagnosis systems also belong) made it necessary to look for roundabout ways. One of these ways being developed intensively at present is the creation of expert medical systems. An expert system is a computer system which incorporates formalized knowledge of specialists in a certain concrete subject and is able to take expert decisions within this subject (to solve problems in such a way as a man-expert would do it).

Efficiency of operation of the expert system depends in the first place on quantity and quality of the information available in its knowledge base. This is a weak point of expert systems because (1) knowledge base is formed on the basis of subjective ideas of experts whose knowledge is limited and (2) specialists are not able to formalize their knowledge as clear rules; moreover, many of them are not aware, on the whole , what rules they follow.

After much money has been spent for development of a great number of various medical expert systems and objective analysis of their efficiency has been carried out, it will appear the same as known a priori, from the definition:

- When solving relatively simple problems of differential diagnosis (those which are successfully solved by specialists using non-formalized approaches), accuracy of diagnosis achieved by means of expert systems and an expert will be close and sufficient.

- But when solving the most important and complicated problems of differential diagnosis, namely when differentiating diseases which are similar as to symptoms; prognosing progress of disease (for instance, acute myocardial infarction complications ), long-term consequences of surgical interference depending on selected type of operation; in case of early diagnosis (including the latent period) of chronic diseases with a prognosis unfavorable for the life (onoclogical diseases, chronic nephrism), etc., accuracy of diagnosis achievable by means of experts systems and expert will be close and substantially insufficient.

2. TAKEN DEFINITIONS AND FORMULATIONS OF DIFFERENTIAL DIAGNOSIS PROBLEMS

Quite a good progress in medical diagnosis and its conversion from intuitive art of a few talented professionals into a strict science with high level of formalization can be achieved only by transfer from the use of subjective diagnosis information provided by experts to objective information concerning dependence of diagnosis on individual characteristics of a patient and symptom values that are generated using methods provided by artificial intelligence.

There is a certain analogy between medical diagnosis and technical diagnosis. Notwithstanding that technical diagnosis problems are substantially less complex, their solution without use of mathematical models on the basis of paradigm adopted in the science at present is considered as not only incorrect but even indecent.

Under artificial intelligence we understand algorithm allowing to construct informative mathematical model of the object (system) under study carrying essential scope of new, untrivial information unknown before about relations between input and output variables of the object under study, on the basis of the table of experimental data describing behavior of the object (system) of any physical nature without additional a priori information about the structure of a model provided by an expert (specialist in this field) using only formalized procedures.

As applied to medical diagnosis, artificial intelligence allows to obtain for each of disease being differentiated a set of differential syndromes inherent only to a certain disease, many of them unknown before.

Under formal differential diagnosis we understand a formalized procedure (including also using computer) enabling to reliably differentiate each disease from the group of diseases similar as to symptoms using appropriate mathematical models.

On the basis of this definition differential diagnosis problems can be as follows:

- differential diagnosis within the group of diseases similar as to symptoms;

- prognosis of disease progress and outcome;

- early diagnosis of diseases dangerous for the life (for instance, cancerous disease in the latent period).

Taking into account the definitions taken the problem of differential diagnosis model construction can be presented as the following mathematical formulation.

GIVEN: Table of experimental data M=XxY (X={Xij}, i=1,m, j=1,n; Y={Yil}, i = 1, m, l=1,k) , each line of which contains information about symptoms values (Xij) and verified diagnosis Yil for the i-th patient. (Here m is a number of lines (patients) in table M, n is a number of columns (symptoms) in table M, K is a number of diseases to be differentiated.)

IT IS NECESSARY: to construct a mathematical model consisting of K of disjunctions of differential syndromes, each disjunction containing differential syndromes of only one of K of diseases to be differentiated on the basis of the table M by means of formalized procedures.

3. SOLUTION METHODS OF FORMAL DIFFERENTIAL DIAGNOSIS PROBLEMS

If a table, each line of which contains symptoms values and verified diagnosis for one patient, has been obtained on the basis of the examination results, and all observed patients belong to a set of diseases difficult to differentiate from each other, being studied, a mathematical model of differential diagnosis can be constructed using a new mathematical simulation method called mosaic portrait.

The essence of the method consists in realization of the following formalized procedures:

a) transformation of initial experimental data table M to table M' using formalized split of range of values of each symptom into subranges ( transfer from continual scales used for measurement of symptoms to discrete ones). In this case a specific code is applied to each subrange of values of each symptom;

b) search of codes combinations which can be found in table M' in lines belonging to one disease and not found in any line with other diseases.

These combinations are interpreted as differential syndromes of a corresponding disease. The mosaic model obtained in such a way consists of K of disjunctions, each containing differential syndromes of one of K of diseases to be differentiated. Number of differential syndromes is very large and most of them are new, untrivial, unknown before in the medical science.

Well-known methods of realization as per paragraph b require complete search for all possible combinations of subranges of values of all symptoms and belong to so called NP-problems. Computer time required for the solution of such problems depends exponentially on the number of input variables.

When the number of symptoms is more than 15 these problems are practically insoluble.

Using the mosaic portrait method, it is possible to solve problems with dimensionality of the order of 1000 input variables for acceptable time period.

4. NEW POSSIBILITIES WHEN USING MOSAIC PORTRAIT METHOD FOR SOLUTION OF MEDICAL PROBLEMS

Use of the mosaic portrait method allows to solve a number of critical problems of medicine solution of which by means of well-known methods caused serious and mostly irresoluble methodical and computing difficulties.

These problems include:

- formal differential diagnosis of diseases similar as to symptoms;

- formal prognosis of disease progress and outcome alternatives;

- formal selection of one case from a set of alternative surgical interference on the basis of criteria of long-term consequences;

- early (or even in the latent period) diagnosis of chronic diseases with the prognosis dangerous for the life (ontological diseases, etc.);

- preliminary screening based on data of mass preventive examinations;

- construction of informative mathematical models using the table of experimental data resulting from any medical examination.

High self-descriptiveness of mosaic models allows to substantially increase the accuracy of differential diagnosis, to substantially reduce its cost and load on a patient by excluding invasive and low-informative diagnostic procedures, to formalize diagnosis procedure and to realize it by computer.

5. EXAMPLES OF PRACTICAL USE OF NEW METHODS OF DIFFERENTIAL DIAGNOSIS

At present there is a large experience in using the mosaic portrait method for construction of models of differential diagnosis of diseases difficult to differentiate and their practical use for formal diagnosis and disease outcome prognosis.

A prognosis method of myocardial infarction complications at the acute phase has been developed in collaboration with Military Medical Academy (Saint-Petersburg) [1,2].

Differential syndromes typical for each of consequences of myocardial infarction have been obtained on the basis of experimental data containing information about patients who died of infarction complications (cardiogenic shock, perfusion insufficiency, ventricular fibrillation, cardiac rupture) and had infarction without consequences. Namely these syndromes were used for diagnosis. Method of myocardial infarction consequences prognosis was adopted at Military Medical Academy (Saint-Petersburg), 23th hospital (Moscow), 20th hospital and 42nd polyclinic (Saint-Petersburg). 80-88% of prognosis have been verified by clinical data and in case of lethal outcome by results of pathologicoanatomic investigation. As a result of preventive treatment based on prognosis data rate of death of large-Indus infarction reduced by 36.8% and of small-Indus infarction by 45.1%.

Based on the results of this work, a software package for realization of an intelligent system and called "prognosis of myocardial infarction complications at acute phase" was developed.

Initial information about the state of a patient including anamnesis, examination data, electric cardiograph data (total 39 indicants) is entered by keystroke in the question-answer mode.

The following problems are solved by this package:

- prognosis of one or several possible myocardial infarction complications according to the information about patient state collected at the first day of his stay in a hospital;

- in case of prognosis of several complications , the probability of realization of each of them;

- every two days - a new prognosis corrected on the basis of treatment results and recommendations on preventive and symptoms-eliminating therapy corresponding to this prognosis;

- on inquiry, displaying of information concerning differential syndromes on which basis the prognosis was done;

- input of information about the patient, complications prognosis, syndromes, on which basis the prognosis was done, and recommended treatment into the data base.

A simple (without gastroscopy) and express method of differential diagnosis "gastric ulcer - gastric cancer" with accuracy of 96.4 % has been developed and put into practice in collaboration with Military and Medical Academy.

Based on the results of this work, a software package for realization of an intelligent system and called "DIFFERENTIAL DIAGNOSIS GASTRIC ULCER - GASTRIC CANCER" was developed.

Method of differential diagnosis of various pathogenesis ("shock lung", aspiration, atelectasis, toxic-septic, hypostatic and bronchial pneumonia ) for burnt patients which ensures early (1-2 days of progress) diagnosis, allows to differentiate the therapy and increase the efficiency of treatment was developed in cooperation with Kharkov Burn Centre [3,4].

On the basis of the results of this work, a software package for realization of intelligent system and called "DIFFERENTIAL DIAGNOSIS of pathogenesis of PNEUMONIA FOR burnt patients" was developed.

Reliable and express methods of differential diagnosis "2nd phase of hypertensive disease - chronic diffuse glomerulonephritis with hypertension" and chronic glomerulo-pyelonephritis" was developed and put into practice in collaboration with Military Medical Academy on the basis of biomicroscopy of bulbar conjunctiva.

6. MAIN TASKS AND FINAL GOAL OF THE WORK

The main task of the work is the creation of an efficient intelligent system having no analogues in the world practice, realized at computer and enabling to carry out correct computer-aided diagnosis within the group of diseases difficult to differentiate using formalized procedures.

The depth of the medical diagnosis system created will be constantly increasing: diagnosis accuracy will become higher due to feedback - model after-education based on reliably verified errors of formal diagnosis.

After development and patenting the intelligent differential diagnosis system will be represented as a final commercial product being PC software package.

As practically all EXPERT SYSTEMS AND INTELLIGENT SYSTEMS well-known as on today, allowing to obtain knowledge using the mosaic portrait method use the same language of algebra of logic, according to which any hypothesis is formulated as proposition "if..... then..." and the expert systems accumulate (or must do it) all well-known knowledge in this field, then in case of expert and intelligent systems created for the same field their intersection (common hypotheses) is an a priori known, trivial information; logic difference between propositions of expert and intelligent systems - misinformation (false information); logic difference between propositions of intelligent and expert systems - new, untrivial information unknown for the specialists in this field.

There is just the reason why intelligent medical systems can replace expert systems in the intelligent medical products market.

7. Program of works

Formal differential diagnosis models can be developed in collaboration with any medical institution (research institute, university, diagnostic center, hospital, etc.):

- specialized in studying (treating) of certain diseases;

- having qualified specialists and up-to-date diagnosis equipment for correct diagnosis verification;

- having archival materials or possibility to examine not less than 60 patient as to each disease of the group of diseases to be differentiated during acceptable time period (up to one year).

7.1 More precise definition of the problem. Required experimental data collection

7.1.1 Finalizing of the list of nosologic units to be differentiated.

7.1.2 Making of the list of input variables which can be used potentially to solve the problem. (This list must be surplus i.e. contain not only well-known variables selection of which is substantiated by positive experience of differential diagnosis of diseases being studied, but also other variables which are included on the basis of intuitive ideas of specialists.

7.1.3 Development of formalized history - unified form containing list of symptoms determined as per paragraph 7.1.2.

7.1.4 Determining of minimum required and sufficient scope of experimental studies to solve the problem of differential diagnosis and choose optimum treatment taking into account individual features of a patient.

7.1.5 Purposeful examination of patients suffering from diseases as per paragraph 7.1.1; organizing of correct verification of diagnosis within the group of diseases under study; recording of results of examination using the unified form (see paragraph 7.1.3).

7.1.6 Collection of experimental data is finished by a table each line of which contains information about symptoms and verified diagnosis for one patient.

7.2 DEVELOPMENT OF INTELLIGENT COMPUTER SYSTEM PROPER CONSISTS OF THE FOLLOWING STAGES:

7.2.1 Construction of mathematical models of differential diagnosis on the basis of table of experimental data (see paragraph 7.1.6).

7.2.2 Minimization of the models as per paragraph 7.2.1.

7.2.3 Evaluation of adequacy of differential diagnosis models using new experimental data.

7.2.4 Repetition of paragraphs 7.2.1 and 7.2.2 using experimental data including the initial table as per paragraph 7.1.6 and the table used to evaluate adequacy of the model (paragraph 7.2.3).

7.2.5 Creation of intelligent medical system for development of software package providing the solution of the problem stated.

7.3 Constant after education of the system

As the errors are not excluded in the diagnosis on the basis of initial experimental data used for construction of a model the errors are also possible in case of formal differential diagnosis. It is found out experimentally that percentage of diagnosis errors in the initial data (accuracy of diagnosis achieved in a medical institution whose experimental data was used) and when using the model agree practically. After education of the model on the basis of additional experimental data recording actual errors of formal diagnosis will enable to construct the next version of the model ensuring diagnosis with less percentage of errors than made by specialists.

7.3.1 Development of formalized procedure of self-improvement of the system-after education of the models using reliably verified errors of formal diagnosis.

7.3.2 Collection of information about actual cases of inaccurate differential diagnosis at medical institutions participating in the development of the system.

7.3.3 Correction of the models and software package using information about actual cases of inaccurate diagnosis. (Sequential development of new, more precise versions of the system).

7.4 PATENTING OF INTELLIGENT MEDICAL SYSTEM

The models obtained by mosaic portrait methods contain a great number of new, unknown in the medical science differential syndromes for each of diseases to be differentiated which will allow to patent the corresponding system.

8. COMMERCIALIZATION OF THE SYSTEM

Arrangement of advertising campaign:

- development of demonstration disks;

- publications in medical journals, papers and presentations of demonstration versions at scientific conferences, delivery of advertising materials to practical medicine institutions, etc..

8.2 Commercialization of the system

References:

1. G.M.Yakovlev, V.N.Ardashev, M.D.Kats, T.A.Galkina. Mosaic portrait method in the myocardial infarction prognostication. Cardiology, 1981, No.6

2. Prognosis of outcome and complications of acute myocardial infarction (edited by V.P.Malygina). Moscow, Voyenizdat, 1987, p.128.

3. L.M.Tsogoeva, D.E.Pekarsky, S.F.Kudrya, M.D.Kats. Mosaic portrait method in differential diagnosis of pneumonia for burned patients. Clinic surgery, 1991, No.3.

4. L.M.Tsogoeva. Differential diagnosis and peculiarities of treatment of various forms of pneumonia for burned patients. Abstract of candidate thesis, Kharkov, 1991.

5. V.S.Zaitsev. Microcirculation state and rheologic properties of blood in case of hypertensive disease, chronic glomerulo- and pyelonephritis. Abstract of candidate thesis. Leningrad, 1984.

Information supplied by the Author December 1998. Page last updated: January 19, 2004