Encrypted code: Ephaptic transmission

Go to top

The term 'Cognitive Neural Network' abbreviated to 'CNN' is a dynamic cognitive intellectual process of the clinician who interrogates the network for self-training. The 'CNN' is not a 'Machine Learning' because while the latter must be trained by the clinician, with statistical and prediction adjustments, the 'CNN' trains the clinician, or rather, directs the clinician to the diagnosis while always being questioned following a logical human, hence the term 'Cognitive'. As demonstrated, the definition of 'Hemasticatory spasm' in our patient Mary Poppins was not a clinically simple process, however, considering the themes, presented in the previous chapters of Masticationpedia, we have at least three supporting elements available: a vision of 'Quantum Probability' of physical-chemical phenomena in the complex biological systems which will be discussed extensively in the specific chapters; a more formal and less vague language than the natural language which directs the diagnostic analysis to the first input node of the 'CNN' through the ' $\tau$ Coherence Demarcator' described in the chapter '1st Clinical case: Hemimasticatory spasm'; the 'CNN' process which, being managed and guided exclusively by the clinician, becomes an essential means for the definitive diagnosis. The 'CNN', in fact, is the result of a profound cognitive process that is performed on each step of the analysis in which the clinician weighs his intuitions, clarifies his doubts, evaluates the reports, considers the contexts and advances, step by step, confronting the result of the answer coming from the database. In our case, the database is Pubmed which substantially represents the current level of basic knowledge $KB_{t}$ at the time of the query and $KB_{c}$ in the broader specialist contexts.

Article by Gianni Frisardi · Flavio Frisardi · Giorgio Cruccu

Introduction

In the chapter '1st Clinical case: Hemimasticatory spasm' we immediately reached a conclusion bypassing all the cognitive, clinical and scientific process which underlies the diagnostic definition but it is not that simple otherwise our poor patient Mary Poppins, would not have had to wait 10 years for the correct diagnosis.

It should be emphasized that it is not a question of negligence on the part of clinicians rather of the complexity of 'biological systems' and above all of a mindset still anchored to a 'classical probability'. The 'Classical probability' categorizes healthy and diseased phenotypes according to symptoms and signs sampled clinicians instead of probing the 'State' of the system in the temporal evolution. This concept, anticipated in the chapter 'Logic of medical language: Introduction to quantum-like probability in the masticatory system' and in 'Conclusions on the status quo in the logic of medical language regarding the masticatory system' has laid the foundations for a medical language more articulated and less deterministic, mainly focused on the 'State' of the 'Mesoscopic System' whose purpose is, essentially, to decrypt the message in machine language generated by the Central Nervous System. We will assist in the description of other clinical cases that will be reported in the next Masticationpedia chapters.

This model, which we propose with the term 'Cognitive Neural Network' abbreviated as 'CNN' is a dynamic cognitive intellectual process of the clinician who interrogates the network for self-training. The 'CNN' is not a 'Machine Learning' because while the latter must be trained by the clinician, with statistical and prediction adjustments, the 'CNN' trains the clinician or rather directs the clinician to the diagnosis while always being questioned following a logical human, hence the term 'Cognitive'.

In fact, some classic machine learning models, whose training in the laboratory gives positive results, fail when applied to the real context. This is typically attributed to a mismatch between the datasets the machine was trained with and the data it encounters in the real world. A practical example of this can be represented by the conflict of assertions encountered in the diagnostic process of our patient Mary Poppins between the dental and neurological context, which only the support of the $\tau$ coherence demarcator (cognitive process) managed to solve.

One of the limits of machine learning, therefore, is known as "data shift",^[1] or "data movement" and another underlying cause of the failure of some models outside the laboratory, is the "subspecification"^[2]^[3] so much so that the attempt to build an algorithm-enhanced electronic medical record (EMR) system, designed specifically for use in a cancer center, was a notable failure at an estimated cost of $39,000,000 USD. This effort was a 2012 partnership between M.D. Anderson Partners and IBM Watson in Houston, Texas.^[4] Early promotional news describing the project stated that the plan was to combine genetic data, pathology reports with physicians' notes and relevant journal articles to help doctors come up with diagnoses and treatments. However, five years later, in February 2017, M.D. Anderson announced that he had closed the project because, after several years of trying, he hadn't produced a tool for use with patients that was ready to move beyond pilot testing.

«Fascinating and provocative, explain to me in detail»
(... the model is essentially simple in its cognitive complexity)

In essence, the encrypted machine language message sent out by the Central Nervous System in the 10 years of illness of our patient Mary Poppins was interpreted, through verbal language, as Orofacial Pain from Temporomandibular Disorders'. We have remarked several times, however, that human verbal language is distorted by vagueness and ambiguity therefore, not being a formal language such as mathematical language, it can generate diagnostic errors. The message in machine language sent out by the Central Nervous System to be searched for is not pain (pain is a verbal language) but the 'Anomaly of System State' in which the organism was in that time period. Hence the shift from the semiotics of the symptom and the clinical sign to the 'System Logic' which, through 'System's Theory' models, quantify the system's responses to incoming stimuli, even in healthy subjects.

All this conceptuality is replicated in the proposed 'CNN' model by dividing the process into incoming triggers (Input) and outgoing data (Output) to then be reiterated in a loop managed cognitively by the clinician up to the generation of a single node useful for the definitive diagnosis. The model basically breaks down as follows:

Input: By incoming trigger, we mean the cognitive process that the clinician implements as a function of the considerations received from previous statements, as has been pointed out in the chapters concerning the 'Medical language logic'. In our case, through the ' $\tau$ $\tau$ Consistency Demarcator, the neurological context was defined as suitable instead of the dental one pursuing a clinical diagnostic explanation of TMDs.
- This trigger is of essential importance because it allows the clinician to point out the network analysis 'Initialization command' which will connect a large sample of data corresponding to the set trigger. To this essential 'Initialization command', as an algorithmic decryption key, is added the last closing command which is equally important as it depends on the intuition of the clinician who will consider the decryption process finished.
- In Figure 1 the structure of the 'CNN' is represented where the difference between the most common neural network structures can be noted and in which the first stage is structured with a high number of input variables. In our 'CNN' the first stage corresponds only to a node and precisely to the network analysis of 'Initialization command' called ' $\tau$ Coherence Demarcator', the subsequent loops of the network, which allow the clinician to terminate or to reiterate the network, (1st loop open, 2st loop open,...... nst loop open) are decisive for concluding the decryption process ( Decrypted Code ). This step will be explained in more detail later in the chapter.

Figure 1:Graphical representation of the 'CNN' proposed by Masticationpedia

Output: The outgoing data from the network, which substantially correspond to a precise cognitive trigger request, returns a large number of data classified and correlated to the requested keyword. The clinician will have to devote time and concentration to continue decrypting the machine code. In fact, we have witnessed how, following the indications dictated by research criteria such as the 'Research Diagnostic Criteria' (RDC), our patient Mary Poppins was immediately categorized as 'TMDs'. We have also suggested a way to expand diagnostic capabilities in dentistry through a 'fuzzy' model that would allow to range in contexts other than one's own. This shows the complexity in making differential diagnoses and the difficulties in following a classical semiotic roadmap because we are anchored too much to verbal language and too little to a quantum culture of biological systems. This borders on the concept of machine language and initialization decryption command that we will briefly explain in the next paragraph.

Initiatialization command

For a moment let's imagine that the brain speaks the language of a computer and not vice versa as happens in engineering, to distinguish the aforementioned difference between machine language and human verbal language. To write a sentence, a word or a formula, the computer does not use the classic verbal mode (alphabet) or the decimal mode (numbers) with which we write mathematical formulas but its own 'writing' language code called html code for the web. Let's take as an example the writing of a fairly complex formula, it is presented to our brain in the verbal language with which we have learned to read a mathematical equation, in the following form:

$+2\sum _{\alpha _{1}<\alpha _{2}}\cos \theta _{\alpha _{1}\alpha _{2}}{\sqrt {P(A=\alpha _{1})P(B=\beta |A=\alpha _{1})}}P(A=\alpha _{2})P(B=\beta |a=\alpha _{2})$ and let us imagine, letting our minds wander, that this formula corresponds to the message of the Central Nervous System, as we have anticipated, in the 'Ephaptic Transmission' still to be decrypted.

The computer and therefore the brain, for our metaphorical example, does not know verbal language or rather it is only a convention generated to simplify natural communication, rather it has its own one with which to write the formula mentioned and in the wiki text language (with extension .php) looks like this, represented in figure 2:

Figura 2: Wiki text of a mathematical formula. Note the initialization <math> command and the script exit </math> command
as we can see it has nothing to do with verbal language and in fact, the brain has its own machine language made up not of vowels, consonants and numbers but of action potentials, wave packets, frequencies and amplitudes, electric dipoles , etc. what we simply observe in an electroencephalographic tracing (EEG) and which represents, precisely, the electromagnetic fields on the scalp of the dipoles's activity and the cerebral ionic currents that propagate in the encephalic volume.

The story, however, does not end here because this is a writing language that has nothing to do with the interpreter of computer hardware and therefore with the organic structure of the brain made up of Centers with specialized functions, synaptic, polysynaptic circuitry and other. This writing language, therefore, derives from a machine language that is not modeled in the command '<math>' rather than '+2\sum_{\alpha_1'} but derives from a binary language subsequently converted into html writing code. This is referred to as 'machine language' for both the computer and the brain and can be simulated as follows

00101011 00110010 01011100 01110011 01110101 01101101 01011111 01111011 01011100 01100001 01101100 01110000 01101000 01100001 01011111 00110001 00111100 01011100 01100001 01101100 01110000 01101000 01100001 01011111 00110010 01111101 01011100 01100011 01101111 01110011 01011100 01110100 01101000 01100101 01110100 01100001 01011111 01111011 01011100 01100001 01101100 01110000 01101000 01100001 01011111 00110001 01011100 01100001 01101100 01110000 01101000 01100001 01011111 00110010 01111101 01011100 01110011 01110001 01110010 01110100 01111011 01010000 00101000 01000001 00111101 01011100 01100001 01101100 01110000 01101000 01100001 01011111 00110001 00101001 01010000 00101000 01000010 00111101 01011100 01100010 01100101 01110100 01100001 01111100 01000001 00111101 01011100 01100001 01101100 01110000 01101000 01100001 01011111 00110001 00101001 01111101 00100000 01010000 00101000 01000001 00111101 01011100 01100001 01101100 01110000 01101000 01100001 01011111 00110010 00101001 00001010 01010000 00101000 01000010 00111101 01011100 01100010 01100101 01110100 01100001 01111100 01100001 00111101 01011100 01100001 01101100 01110000 01101000 01100001 01011111 00110010 00101001

But what if the following string 00101011 00110010 01011100 01110011 01110101 01101101 which corresponds to the <math> command is not present in this code?

The message would be corrupted and the formula would not be generated due to lack of the most important step that of 'Initialization of the command code', as well as if we eliminated the last part of the code 01101100 01110000 01101000 01100001 01011111 00110010 00101001, corresponding to the closure of the script < /math> the formula would remain corrupted and indeterminate.

In practice, without the initial and final command, the formula is well described in the following form that is understandable to us:

$+2\sum _{\alpha _{1}<\alpha _{2}}\cos \theta _{\alpha _{1}\alpha _{2}}{\sqrt {P(A=\alpha _{1})P(B=\beta |A=\alpha _{1})}}P(A=\alpha _{2})P(B=\beta |a=\alpha _{2})$

it would present itself in a way incomprehensible to most people.

+2\sum_{\alpha_1<\alpha_2}\cos\theta_{\alpha_1\alpha_2}\sqrt{P(A=\alpha_1)P(B=\beta|A=\alpha_1)} P(A=\alpha_2) P(B=\beta|a=\alpha_2)

Just as the lack of part of the binary code corrupts the representation of the formula, similarly the decryption of the machine language of the CNS is a source of vagueness and ambiguity of the verbal language and contextually of diagnostic error.

Cognitive process

The heart of the 'CNN' model lies in the cognitive process referred exclusively to the clinician who is at the helm while the network essentially remains the compass that warns of off course and/or suggests other alternative routes but the decision-making responsibility always refers to the clinician ( human mind). In this simple definition, we will perceive it better at the end of the chapter, the synergism 'Neural network' and 'Human cognitive process' of the clinician will be self-implementing because, on the one hand, the clinician is trained or better guided by the neural network (database), while, the last one will be trained on the latest updated scientific-clinical event. Basically, the definitive diagnosis will add an additional piece of information to the temporal base knowledge $Kb_{t}$ . This model differs substantially from 'machine learning' just by observing the two models in their structural configuration (Figures 1 and 3).

Figure 3: Graphic representation of an archetypal ANN in which it can be seen in the first stage of initialization where there are five input nodes^[5] while in the 'CNN' model the first stage is composed of only one node. Follow text.

Figure 3 shows a typical neural network, also known as artificial NNs. These artificial NNs attempt to use multiple layers of calculations to mimic the concept of how the human brain interprets and draws conclusions from information.^[5] NNs are essentially mathematical models designed to handle complex and disparate information, and this algorithm's nomenclature comes from its use of synapse-like "nodes" in the brain.^[6] The learning process of a NN can be supervised or unsupervised. A neural network is said to learn in a supervised manner if the desired output is already targeted and introduced into the network by data training, while, unsupervised NN has no such pre-identified target outputs and the goal is to group similar units close together in certain areas of the range of values. The supervised module takes data (e.g., symptoms, risk factors, imaging and laboratory findings) for training on known outcomes and searches for different combinations to find the most predictive combination of variables. NN assigns more or less weight to certain combinations of nodes to optimize the predictive performance of the trained model.^[7]

Figure 1, on the other hand, corresponds to the 'CNN' model proposed and it can be seen how the first stage of acquisition is composed of a single node while the 'Machine learning' at the first node, the more incoming variables have the greater the 'Prediction' in exit. As mentioned, it should be taken into account that the first node is of fundamental importance because it already derives from a clinical cognitive process that led the ' $\tau$ Coherence Demarcator' to determine a first-ever choice of field. From the 'Initialization command', therefore, the neural network evolves in a series of states composed of a large number of nodes and then terminates at a first step of one or two nodes and then reiterates in a subsequent loop of several nodes until ending in the 'last conclusive node (decryption of the code). The initialization process of the first node, the last and the reiteration of the loop is exclusive to the human cognitive process of the clinician and not to a statistical automatism of machine learning, much less 'Hidden' stages. All open and closed loops must be known to the clinician.

For further information on the subject, it is available on Masticationpedia in the chapter 'An artificially intelligent (or algorithm-enhanced) electronic medical record in orofacial pain'

But let's see in detail how a 'CNN' is built

Cognitive Neural Network

In this paragraph it seems necessary to explain the clinical process followed with the support of the 'CNN' following step by step the cognitive queries to the network and the cognitive analysis performed on the data in response from the network. The map has also been shown in figure 4 with links to the network responses that can be viewed for more consistent documentation:

$\tau$ Coherence Demarcator: As we have previously described, the first step is a network analysis initialization command that derives, in fact, from a previous cognitive processing of the assertions in the dental context $\delta _{n}$ and the neurological one $\gamma _{n}$ to which the ' $\tau$ Coherence Demarcator' gave absolute weight by, effectively, eliminating the dental context $\delta _{n}$ from the process. From what emerges from the neurological assertions $\gamma _{n}$ the 'State' of the Trigeminal Nervous System appears unstructured highlighting anomalies of the trigeminal reflexes for which the 'Initialization' command will be the 'Trigeminal Reflex' to go and test the database (Pubmed).
1^st loop open: This 'Initialization command', therefore, is considered as initial input for the Pubmed database which responds with 2,466 clinical and experimental data available to the clinician. The opening of the first true cognitive analysis is elaborated precisely on the analysis of the first result of the 'CNN' corresponding to 'Trigeminal Reflex'. In this phase we realize that a discrete percentage of data reveal a correspondence between trigeminal reflex abnormalities and demyelination problems, therefore the 1^st loop open will correspond to: 'Demyelinating neuropathy'' which will return 14 sensitive data. Behind the choice of this key there is an active and dynamic cognitive process of the clinician. From the assertions in the neurological context, a neuropathic pathology was hypothesized in which the demyelinating aspect should also be considered.
2^st loop open: The process continues by focusing in ever greater detail on the keywords that correspond to our electrophysiological anomalous result data, i.e. the latency of the jaw jerk. This input corresponding to 'Latency' returns 6 sensitive data on which to process a further iteration of the loop.
3^st loop open: In the statements of the neurological context, an anomaly is also observed in the amplitude of the jaw jerk as well as in latency. This 3^st open loop corresponding to 'Amplitude' returns only 2 data points on which to dwell to decide the keyword to reiterate the loop or definitively close the process. The result shows an article describing the electrophysiological evaluation of cranial neuropathies that was considered of low specific weight for our purposes while the other article highlights some trigeminal methodologies to test latency, amplitude of masticatory muscles including H-reflex .
4^st loop closed: The process, therefore, continues by inserting the algorithmic keyword 'H-reflex ' which returns 3,701 clinical scientific data.
5^st loop open: The anomalies highlighted were mainly verified on the masseters so it can be deduced that keywords concerning the 'Masseter muscle' can be intercepted in the interrogated sample of the 4^st closed loop, hence the 5^st open loop which returns 30 data available for the 'CNN'
6^st loop open: We, however, do not know whether the neuropathic damage is localized exclusively on the masseter muscle or also involves the temporal muscle, therefore another algorithmic keyword would be the 'Temporal muscle' which returns 8 sensitive data.
7^st loop open: From a careful analysis of this 7th open loop one wonders if these electrophysiological anomalies can be highlighted in patients with sclerosis and being present in the patient's clinical history, a previous diagnosis of 'Morphea' it was opted to interrogate the 'Network' of a another keyword focused on 'Sclerosis' which gave only one sensitive data 'Heteronymous H reflex in temporal muscle as sign of hyperexcitability in ALS patients
8^st loop closed: In this single node the clinician could terminate the loop but would not have solved anything because the decoding of the encrypted message has not yet been achieved. It should be noted that the electrophysiological method called 'heteronomous H-reflex' is able to highlight response anomalies from the temporal muscle for which the loop was continued by inserting the following specific keyword, ' Temporal muscle abnormal response' which returns 137 data.
9^st loop open: By studying the 137 articles appeared in Pubmed, the clinician intuits that the response abnormalities, in the temporal muscle through the H-reflex test, depend on a spread of the stimulus current in the unstructured axon and therefore further investigates the loop by interrogating the network for a further keyword the 'Lateral spread impulses' which, definitively, closes the cognitive process of the 'Neural Network' with a close article, the 'Ephaptic transmission is the origin of the abnormal muscle response seen in hemifacial spasm'

Figure 4: The active links of the 'CNN' highlighted in the template correspond to the 'Pubmed' database and can be documented.

Conclusion

As demonstrated, the definition of Hemasticatory spasm in our patient Mary Poppins was not a clinically simple process, however, considering the topics presented in the previous chapters of Masticationpedia we have at least three elements of support available:

A vision of 'Quantum Probability' of physical-chemical phenomena in complex biological systems which will be discussed extensively in the specific chapters.
A more formal and less vague language than the natural language that directs the diagnostic analysis to the first input node of the 'CNN' through the ' $\tau$ Coherence Demarcator' described in the chapter '1st Clinical case: Hemimasticatory spasm'
The 'CNN' process which, being managed and guided exclusively by the clinician, becomes an essential means for the definitive diagnosis.

The 'CNN', in fact, is the result of a profound cognitive process that is performed on each step of the analysis in which, the clinician weighs his intuitions, clarifies his doubts, evaluates the reports, considers the contexts and advances step by step confronting the result of the answer coming from the Pubmed database. Substantially the Pubmed database represents the current level of basic knowledge $KB_{t}$ at question time and the $KB_{c}$ in the broadest specialist contexts.

A linear representation of this cognitive process labeled as $CNN_{1}$ with the necessary annotations it could be the following:

$CNN_{1}=\sum ($ Trigeminal Reflex,Demyelinating neuropathy, Latency,Amplitude,H-reflex, Masseter muscle, Temporal muscle, Heteronymous H reflex in temporal muscle as sign of hyperexcitability in ALS patients,Temporal muscle abnormal response $)$ $\longrightarrow$ Ephaptic transmission is the origin of the abnormal muscle response seen in hemifacial spasm

There are essentially two annotations to observe: the first is the mandatory identification of the 'Initialization command' that derives from the context chosen through the ' $\tau$ Coherence Demarcator' and the second the order of the cognitively selected keywords.

«In fact, what happens to the 'CNN' if we consider the dental context by inserting the keyword 'Temporomandibular Disorders' as an initialization input, keeping everything else unchanged?»

By marking the network as $CNN_{2}$ with a dental initialization input ( Temporomandibular disorders) as follows:

$CNN_{2}=\sum ($ Temporomandibular Disorders,Trigeminal reflex, Demyelinaying neuropathy, Latency $\longrightarrow$ Side asymmetry of the jaw jerk in human craniomandibular dysfunction

The message is corrupted, as explained above regarding the mathematical formula, as the 'Initialization command input' ( Tempormandibular disorders) directs the network for a set of data, no less than 20,514, which lose connections with a part of subsets. While maintaining the rest of the CNN similar to the previous one (neurological context) the network stops at the keyword 'latency'. In fact, only one scientific article was extracted and obviously concerns the latency of the jaw jerk but is not related to neuropathic disorders. (figure 5) The error, in the choice of the initialization command input of the $CNN$ process not only corrupts the message to be decrypted but renders vain all the upstream work of analysis of the clinical assertions discussed in the chapters of language logic.

Figure 5: Ending the 'CNN' loop with an initialization input of 'Tempororomandibular Disorders'

However, changing the order of the keywords in an exact cognitive path such as the neurological one, essentially, returns the same results as the previous one provided that the initialization command input is perfectly focalized, as can be seen in the following simulation labeled with $CNN_{3}$ :

$CNN_{3}=\sum ($ Trigeminal reflex, amplitude latency demyelinating neuropathy $\longrightarrow$ H-reflex................

and reconnects to the previous one until closing in the output 'Ephaptic transmission is the origin of the abnormal muscle response seen in hemifacial spasm' (Figure 6)

Figure 6: Process of 'CNN' with order of keywords changed.

Note that if this chapter had been published in an impacted international scientific journal (Impact Factor) the $KB_{t}$ and contextually a hypothetical 'machine learning' would have been enriched with a new content or that of the diagnosis of 'Hemimasticatory spasm' defined following the electrophysiological method of the heteronymous H-reflex. This conclusion will come in handy when we repeat the same procedure for other clinical cases in which the $KB_{t}$ is updated to the database output.

Bibliography

↑ Jérôme Dockès, Gaël Varoquaux, Jean-Baptiste Poline. Preventing dataset shift from breaking machine-learning biomarkers.GigaScience, Volume 10, Issue 9, September 2021, giab055,
↑ Alexander D’Amour et al. Underspecification Presents Challenges for Credibility in Modern Machine Learning. Journal of Machine Learning Research 23 (2022) 1-61,Submitted 11/20; Revised 12/21; Published 08/22
↑ Damien Teney, Maxime Peyrard, Ehsan Abbasnejad. Predicting Is Not Understanding: Recognizing and Addressing Underspecification in Machine Learning.ECCV 2022: Computer Vision – ECCV 2022 pp 458–476Cite as
↑ Herper M. MD Anderson benches IBM Watson in setback for artificial intelligence in medicine. Forbes. 2017 February 19. [Ref list]
↑ ^5.0 ^5.1 G S Handelman, H K Kok, R V Chandra, A H Razavi, M J Lee, H Asadi. eDoctor: machine learning and the future of medicine.J Intern Med.2018 Dec;284(6):603-619.doi: 10.1111/joim.12822. Epub 2018 Sep 3.
↑ Schwarzer G, Vach W, Schumacher M. On the misuses of artificial neural networks for prognostic and diagnostic classification in oncology. Stat Med 2000; 19: 541–61.
↑ Abdi H. A neural network primer. J Biol Syst 1994; 02: 247–81.

[1] Jérôme Dockès, Gaël Varoquaux, Jean-Baptiste Poline. Preventing dataset shift from breaking machine-learning biomarkers.GigaScience, Volume 10, Issue 9, September 2021, giab055,

[2] Alexander D’Amour et al. Underspecification Presents Challenges for Credibility in Modern Machine Learning. Journal of Machine Learning Research 23 (2022) 1-61,Submitted 11/20; Revised 12/21; Published 08/22

[3] Damien Teney, Maxime Peyrard, Ehsan Abbasnejad. Predicting Is Not Understanding: Recognizing and Addressing Underspecification in Machine Learning.ECCV 2022: Computer Vision – ECCV 2022 pp 458–476Cite as

[4] Herper M. MD Anderson benches IBM Watson in setback for artificial intelligence in medicine. Forbes. 2017 February 19. [Ref list]

[:1-5] 5.0 ^5.1 G S Handelman, H K Kok, R V Chandra, A H Razavi, M J Lee, H Asadi. eDoctor: machine learning and the future of medicine.J Intern Med.2018 Dec;284(6):603-619.doi: 10.1111/joim.12822. Epub 2018 Sep 3.

[6] Schwarzer G, Vach W, Schumacher M. On the misuses of artificial neural networks for prognostic and diagnostic classification in oncology. Stat Med 2000; 19: 541–61.

[7] Abdi H. A neural network primer. J Biol Syst 1994; 02: 247–81.

[1]

[2]

[3]

[4]

[5]

[6]

[7]