Medical Diagnostics and Artificial Intelligence; Diagnostic Constraints
Abstract
Medical diagnostics consists of a selection of various investigative methodologies. For instance: imaging [MRI, CT, Ultrasound, etc.], Electronic/electrical [neural, cardiology], biochemical [disease markers, product of pathological conditions], thermodynamics [fever; viral resp. bacterial], resp. Chemical [cellular communication, poison, drug -prescription / recreational], optical [eye] and mechanical. A variety of combinations of examination factors are generally under consideration for a complete and in-depth analysis. The diagnosis should offer a root-cause analysis, not isolated symptoms.
Introduction
Accurate diagnostics and patient care generally and mandatorily rely on the combined anatomical, physiological and biochemical information obtained from various sources and under the use of many different (complementing) mechanisms-of-action. The acquisition of as much relevant information as possible, and subsequent automated processing will allow for an accurate and relevant diagnosis, followed by the definition of a reliable treatment protocol. The access to the patient’s own medical history files will be crucial in the diagnosis. Furthermore, the comparison against known pathological patterns of other patients will provide screening options to assess the patient health conditions early and determine the most appropriate treatment modalities or determine the need for additional follow-up test in other physiological, anatomical, or biochemical matters or body specific locations. This is just as important as placing the respective patient in the appropriate segment of the statistical distribution with respect to the perceived pathological framework.
Means and Methods
Another important assessment functionality is the risk stratification belonging to the patient conditions and the required urgency of the palliative and curative care.
One specific example of the limitations of single source screening is the use of x-ray imaging only to assess coronary occlusions, and the determination of risk-factors of developing life-threatening cardiovascular conditions, and at which rate of development. One aspect of considerable importance associated with cardiac health is the development of conditions leading up to heart-failure and imminent death.
One important factor, frequently overlooked in diagnostic evaluation, is the medication portfolio associated with the patient’s health profile. Drugs, both recreative and prescribed, can affect the automated (AI-based) first impression created by chosen diagnostic modality. Having access to the full patient history will allow for a well-informed decision process. Comparing the diagnostic data to a database of patients presumably residing in a similar medical category will allow for a quick ranking of risks and assigning an appropriate treatment modus operandi. More importantly, pointing out the level of confidence with which the diagnostic can be made and recommending additional investigative actions. Supporting multi-model diagnostics includes various mechanisms-of-action. For instance, a large portion of the chemical conditions of a patient can nowadays be assessed with spectral imaging, even using a mobile phone.
Since laboratory access may be limited due to geographical location, or the restrictions to local investment in healthcare. Cheap and interlaced diagnostic modalities with over-the-internet evaluation capabilities can bring advanced, accurate and appropriate healthcare to remote and undervalued parts of the world. More importantly, interconnected diagnostics will raise the value of appraising treatment options in a busy, time-limited modern metropolitan healthcare facility as well. Root-Cause analysis may involve investigating whether the pathology is the result of either a viral or bacterial infection, respectively poisons, or recreational or prescription drugs, next to cellular aberrations (e.g. cancer) with potentially genetic foundation.
Results
Artificial Intelligent (AI) generally cannot answer the question: “what is wrong with this patient”. AI often assists in routine analyses of images, respectively signals when there is a suspicion about a specific ailment, such as for instance Breast cancer. When relying on AI to perform the calculations of the physiological value of diagnostic interest, the error (i.e. standard deviation) with respect to each parameter in the function defining the final clinical parameter is carried forward. Remember, calibration of the instrumentation used in the data acquisition with respect to the clinical assessment of the patient does not improve the accuracy of the determinant quantity, it only adjusts for the boundary conditions. These boundary conditions may include the drift in electronic configuration and electronic values (e.g. resistance, capacitance), as well as climatological conditions. The acquired images or data is distrusted over the range of errors in the following manner, described by the following four types of distribution functions: Uniform distribution; Binomial distribution; Bernoulli distribution; Poisson distribution. Normal distribution for instance has the following reliability intervals: 68%, 95%, or 99.7%. Under these conditions 68% of data resides in the first standard deviation range centred on the mean value of the data, 95% of data spans the second standard deviation range, and 99.7% of data reaches out to the edge of third standard deviation range. This means only 68% of patient is within +/-17% of the proverbial “mean”, while the remaining 32% of patients has pathological expressions which are “not normal”, or rather are deviating from a routine protocol and will not, under AI analysis, provide the required accuracy of determination of the illness for at least 32% of patients.
Conclusions
AI generally can only confirm, respectively discard the suspicion of a physician. AI can provide probability distribution of potential causes and factors defining the illnesses of a patient, not accounting for the personal medical history, respectively personal health condition, i.e. athlete or habitual drinker and couch-potato etc. Any researcher, for that matter, any and every physician should consider the possibilities that a patient presents symptoms which are not “the norm” for a potentially serious affliction. Most physiological parameters are a function of age, gender, physical activity history and foremost genetical heritage / chronicle. However, this does not play-out in an identical manner for each individual. Most often the phenomenological expression of a disease [pathological condition] may not be identical for everyone. The advantage of AI in medical diagnosis is however the time-saving aspect, relieving the physicians of a significant burden in the analysis of the conditions of the patients. The physicians still need to keep the broad range of conditions with respect to the various physical conditions of the specific patient in focus for the ultimate diagnosis, always having reservations about the assessments made by AI.
