Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 32 No. 3 (2025) 

 

565 https://internationalpubls.com 

Enhancing the Medical Diagnosis System and Treatment by 

Counterfactual Diagnostic Algorithm 

 

Dr. K. Rama Krishna1, P Prabakaran2, Prof. Monali Shetty3, Vikash Sawan4, Dr. Hari 

Jyothula 5, Dr.S.Suma Christal Mary Sundararajan 6 
1Department of Information Technology, Vasavi College of Engineering, Hyderabad, India. 

k.ramakrishna@staff.vce.ac.in 

2Department of Computer Science and Engineering, Koneru Lakshmaiah Education Foundation, Vaddeswaram, India. 

pprabakaran@kluniversity.in 

3Department of Computer Engineering, Fr. Conceição Rodrigues College of Engineering, Mumbai,India. 

shettymonalin@gmail.com 

4Department of Computer Engineering & Applications, GLA University, Mathura, Uttar Pradesh, India. 

vikash.sawan@gla.ac.in 

5 Department of Computer Science and Engineering, Aditya University, Surampalem, India. dr.jyothulahari@gmail.com 

6 Department of Information Technology, Panimalar Engineering College, Chennai, India. sumasheyalin@gail.com 

 

Article History: 

Received: 03-08-2024 

Revised: 24-09-2024 

Accepted: 07-10-2024 

Abstract:  

Clinical diagnosis and decision-making stand to be radically altered by machine 

learning. By identifying the underlying medical conditions, doctors may better explain 

their patients' symptoms via medical diagnosis. There is a substantial correlation 

between a patient's symptoms and medical history, yet current diagnostic algorithms 

only detect illnesses that are associated with each other. Suboptimal or even harmful 

diagnoses may arise from this failure to differentiate correlation from cause, as this 

article demonstrates. To get around this, they create novel algorithms for 

counterfactual diagnostics and reframe diagnosis as a work of counterfactual inference. 

In this work, researchers demonstrate that this method greatly enhances the reliability 

and security of the subsequent diagnosis while getting closer to the way doctors think 

about health problems. Using a battery of clinical scenarios, they evaluate the 

counterfactual method against 45 medical professionals, the gold standard Bayesian 

diagnosis system, and other algorithms. Compared to the counterfactual approach, 

which reaches expert clinical accuracy, the Bayesian algorithm obtains correctness, 

ranking in the top 26% of clinicians in the study unit. Changing this querying strategy 

alone yields this improvement, eliminating the need for any further model 

enhancements. Based on these results, the use of machine learning in medical diagnosis 

is incomplete without counterfactual thinking. 

Keywords: Decision Making, Machine Learning, Counterfactual Diagnosis, Health 

Problems, Clinical Accuracy. 

      

1. Introduction 

The fundamental problem for healthcare systems worldwide is to provide an accessible and accurate 

diagnosis. Misdiagnosis occurs in around 4% of outpatient visits in India annually. Misdiagnoses of 

major medical illnesses affect a disproportionately large number of people; one-third of these cases 

mailto:shettymonalin@gmail.com
mailto:vikash.sawan@gla.ac.in


Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 32 No. 3 (2025) 

 

566 https://internationalpubls.com 

lead to severe patient injury, and around 21% of these individuals get an incorrect diagnosis at the 

primary care level. 

Machine learning and artificial intelligence have recently grown into potent resources for tackling 

complicated challenges across many fields. Machine learning-aided diagnosis in particular has great 

potential to transform healthcare by making use of vast amounts of information about patients to 

provide accurate and tailored diagnoses. In differential diagnosis, when a patient's symptoms might 

have numerous origins, diagnostic algorithms have failed to match the accuracy of human physicians, 

despite resurgent business interest and extensive research. Because of this, it's important to know why 

current methods have such a hard time with differential diagnosis. By analysing the correlation 

between a patient's symptoms and medical history, all current diagnostic algorithms, including Deep 

Learning and Bayesian model-based methods, use associative inference to diagnose illnesses. On the 

other hand, while making a diagnosis, clinicians often look for conditions that would best explain the 

patient's symptoms. Among the several inference systems that have been identified, associative 

inference stands out as the most basic. At the very top of this hierarchy is counterfactual inference, 

which enables one to assign causal explanations to facts. The main point is that making a diagnosis 

boils down to a game of counterfactual inference. The inability to separate correlation from causation 

severely limits the precision of associated diagnostic algorithms, which can lead to less-than-ideal or 

even harmful diagnoses. To overcome this, the authors suggest a causal concept of diagnosis that is 

more in line with how clinicians make decisions and further get fresh counterfactual diagnostic 

methods to support this method. 

Utilising an experimental set of 1672 clinical vignettes, they evaluate the accuracy level of the 

proposed counterfactual algorithms against a cohort of 45 clinicians and a cutting-edge associative 

diagnostic algorithm. The algorithm which is in associative places in the topmost 49% of physicians 

in the study cohort with an accuracy of 72.53%, compared to the doctors' average diagnosis accuracy 

of 71.41% in the studies. In contrast, the counterfactual algorithm attains expert clinical accuracy and 

ranks in the highest 26% of the cohort with an average accuracy rating of 77.27%. Because diagnostic 

mistakes are more prevalent and can have more severe consequences for uncommon illnesses, these 

advancements stand out. When comparing the associative method with the counterfactual approach, 

they find that the latter provides much higher diagnosis accuracy for uncommon and very rare 

disorders. 

The important thing is that both the associative and counterfactual algorithms use the same illness 

model; the only difference is the querying mechanism. This means that both algorithms improve upon 

each other. Because learning illness models consumes a lot of resources, backwards compatibility is 

crucial. Current Bayesian diagnostic models, whether in or out of the medical field, may therefore 

benefit from the techniques as an instant improvement. 

2. Diagnosis 

The guiding ideas and presumptions of the present algorithmic diagnostic methodology are outlined 

below. As a bonus, they detail the scenarios where causal confounding causes this technique to fail 

and provide a set of guidelines for creating diagnostic procedures that avoid these problems. Lastly, 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 32 No. 3 (2025) 

 

567 https://internationalpubls.com 

let's suggest two novel diagnostic algorithms that are founded on the ideas of adequate and necessary 

causality, using these principles as a basis.  

2.1 Associative Diagnosis 

The use of a method θ to predict the probability of a fault element D given results ε has been 

associated with a model-based diagnosis from its official description. 

𝑷(𝑫|𝜺; 𝛉)                                        (𝟏) 

In medical terminology, D stands for a disease or disease, while ε may refer to symptoms, test results, 

and pertinent medical history. Potential illnesses are rated according to their likelihood when 

diagnosing over several possibilities, as in a differential diagnosis. There are two main types of model-

based diagnostic algorithms: those that discriminate, which model the conditional probability 

distribution for diseases D given input characteristics E (1), and those that generate, which model the 

previous distribution of illnesses and results and calculate the posterior using Bayes rule. 

𝑷(𝑫|𝜺; 𝜽) =
𝑷(𝑫|𝜺; 𝛉)𝑷(𝑫; 𝛉)

𝑷(𝜺; 𝛉)
               (𝟐) 

Generative models are usually Bayesian networks, while discriminative diagnostic models are 

examples of deep learning and neural networks.  

The methods used in this technique are similar to the way physicians make diagnoses. The search for 

underlying causes of a patient's symptoms has long been considered an essential part of the diagnostic 

procedure. In this case, for instance, Habitual abduction is the process by which hypotheses are 

generated. A doctor's ability to diagnose a patient's symptoms is dependent on her familiarity with 

potential reasons. Similarly, a diagnostic is defined as an examination of the source or origin of an 

issue, ailment, or problem. In other words, a doctor will try to diagnose the patient's symptoms by 

looking at the patient's history and current health status to identify the illnesses that are most likely to 

be causing those symptoms. The following is a proposed causal definition of diagnosis: 

• The process of determining, using a patient's medical history, which illnesses are most likely 

to be causing their symptoms. 

There are a lot of literary works that say causal reasoning should be at the heart of diagnosis, yet as far 

as they are aware, no model-based diagnostic approaches use contemporary causal analysis methods. 

In most causal contexts, employing the posterior to determine causal linkages might result in false 

findings due to confounding. Figure 1 (a) displays a condition D that causes symptom S. In this case, 

addressing D may relieve S. Figure 1 (b) shows that variable R complicates the relationship between 

S and D; for instance, R may be a hereditary factor that raises a patient's risk of getting illness D and 

suffering symptom S. In this case, there is a substantial correlation between D and S; nonetheless, a 

credible diagnosis would not be possible if P and D were to have created symptom S. It is not possible 

to easily untangle the directed and common sources of symptoms in disorders, as seen in figure 1 (c). 

Since the posterior cannot distinguish between these many situations, it is inadequate for diagnosing 

symptoms from patients in all but the most basic conditions, particularly in cases when there are several 

potential reasons. 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 32 No. 3 (2025) 

 

568 https://internationalpubls.com 

 

Figure 1 a) Symptom S is caused by disease D; b) D is not the primary source of symptom S, but there 

is a latent common factor R that correlates them; and c) Both disease D and the presence of the latent 

similar cause R produce symptom S. 

2.2 Guidelines for Making Diagnoses 

As an alternate method to associative diagnosis, one might consider the likelihood of causal 

responsibility or the likelihood that the effect S was caused by the target cause D. This calls for a 

diagnostic metric M(Dε) to assess the probability that an illness D is the cause of a patient's symptoms 

based on evidence E. Three very minimum requirements for any such diagnostic tool: 

• Consistently, the chance of a disease D eliciting signs should be proportionate to its posterior 

likelihood (M(Dε) P (D =T ε)). 

• No diagnosis can be made for a disease D that does not induce any symptoms, since M(Dε) 

=0 (causality). 

• Simpler diseases that explain more patient symptoms are more probable. 

Desiderata is justified as follows. According to Desideratum i), the chance of an illness causing 

symptoms is proportionate to the possibility of the patient having the condition. If illness D cannot 

root any of the symptoms from the patients, then D should be ignored (Desideratum ii). Desideratum 

iii) relies on Occam's razor approach, supporting humble diagnosis with few illnesses to explain 

symptoms. The posterior only meets the first desiderata, breaching the final two. 

2.3 Counterfactual Diagnosis 

Utilising counterfactual inference, researchers may measure the probability that a patient's symptoms 

are caused by a disease. Using a counterfactual, one may determine whether an alternative set of 

circumstances would have led to a different result. With the given data E = e, they may determine the 

probability that, in the absence of some hypothetical intervention, they would have seen a different 

result-𝛆 = 𝒆′ , which runs contrary to the fact that ε = e. In Pearl's calculus of interventions, the 

intervention which sets the parameter rX to the value X = x is denoted by do (X = x) Hence the  

counterfactual probability is expressed as 𝑷(𝛆 = 𝒆′|𝛆 = 𝐞, do (X = x)). Counterfactuals quantify the 

effectiveness of an ailment the hypothesis D = T in explaining symptoms evidence S = T by calculating 

the probability that the symptom is contemporary if the disease were cured (P (S = F S = T, do (D = 

F)) (do (D = F)). If the likelihood is higher, D = T provides a valid fundamental explanation for the 

symptom. This probability cannot be expressed as a normal posterior since it pertains to two 

contradicting S states. Appendix C, explain the calculation of counterfactual probabilities. 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 32 No. 3 (2025) 

 

569 https://internationalpubls.com 

The circumstances of this analysis served as inspiration for the two proposed counterfactual diagnostic 

metrics: anticipated sufficiency and expected disablement. Both measures meet all three requirements 

from section 2.2 as shown in the first theorem at the conclusion of this section. 

First Definition (Predictable Impairment): The predicted impairment of illness D is the range of 

symptoms that would change if the disease were cured. 

ᴇ𝒅𝒊𝒔(𝑫, 𝛆) = ∑ |𝑺+

𝒔′

\𝑺+
′ |𝑷(𝒔′|𝛆, 𝐝𝐨 (𝐃 = 𝐅))             (𝟑) 

In this case, E represents factual proof and S+ represents positively proven symptoms. Summarise all 

counterfactual symptom evidence statements. S and S+ represent counterfactual symptoms that are 

affirmatively demonstrated. do (D = F) represents 𝐃 → 𝐅 counterfactual intervention. The cardinality 

of the set of symptoms that are factually present but not in the counterfactual symptoms data is denoted 

by |𝐒+\𝐒+
′ |. 

The concept of required cause is where the predicted disablement is derived from. If S = T, then D is 

a required source of S and vice versa. Consequently, the anticipated disability measures both the 

adequacy of illness D in explaining the symptoms from patients and the probability that treatment D 

alone would ameliorate those symptoms. 

Second Definition (Adequate): D Disease sufficiency is the number of positive symptoms predicted 

to exist after excluding all other plausible causes. 

ᴇ𝒔𝒖𝒇𝒇(𝑫, 𝛆) ≔ ∑ 𝑺+
′ |𝑷(𝒔′|

𝒔′

𝛆, 𝐝𝐨 (𝐏𝐚 (𝑺+)\𝑫 = 𝑭))            (𝟒) 

Where the aggregate of all potential confounding symptoms indicates in the hypothetical symptom 

condition, the symptoms that are positively demonstrated are denoted by 𝒔′ 𝒂𝒏𝒅 𝒔′. In this context, 

Pa(S+) D refers to the set of all directly causative variables for the set of affirmatively proved 

symptoms that do not include illness D, and do 𝑷𝒂(𝑺+)\𝑫stands for the counterfactual interference 

that sets all 𝑷𝒂(𝑺+
′ \𝑫) → 𝑭. 𝛆stands for the collection of all evidence based on facts. In the case of 

hypothetical symptoms, |𝑺+
′ | indicates the cardinality of the collection of all possible symptoms.  

The expectation of sufficiency is based on the concept of adequate cause, which states that if the 

existence of D suggests the later occurrence of S, then D is an adequate cause of S. Nevertheless, the 

anticipated sufficiency is not met since S could exist for various causes; that is, S's presence does not 

automatically indicate that D occurred before. In most cases, symptoms may be adequately explained 

by illnesses. Following the execution of counterfactual actions, every possible reason for the symptoms 

has been eliminated, leaving just D as a potential cause. They next isolate its impact as an adequate 

reason in the theoretical framework. 

Theorem 1: qualities of predicted disability and expected sufficiency as diagnostic criteria. The three 

requirements from section 2.2 are met by anticipated incapacity and anticipated sufficiency. 

 

 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 32 No. 3 (2025) 

 

570 https://internationalpubls.com 

3. Methodology 

In this part, the developers provide the statistical illness models that will be used to evaluate the 

diagnostic metrics discussed before. Simpler formulas for the predicted sufficiency and disablement 

in these models are subsequently derived. 

3.1 Diagnostic Structural Causal Models 

Bayesian Nets represent the links between hundreds of risk factors, illnesses and symptoms; 

researchers apply these models to the study. BNs are often used in diagnostics because they can be 

understood and specifically represent the relationships between variables that are necessary for causal 

and hypothetical analysis. In these models, illnesses, symptoms, and risk factors are usually shown as 

true or false binary nodes. The conventional integer notation is used to represent true as 1 and false as 

0, respectively. 

Directional acyclic graphs (DAGs) and joint probability distributions across all nodes that factorise 

about the DAG architecture define BNs. Assuming a fixed arrow exists between two nodes X and Y, 

say that X is a parent of Y and that Y is a child of X. Assuming a directed link exists between two 

nodes Y and Z, say that Z is an ancestor of Y. Figure 2 (a) shows a basic example BN that models’ 

illnesses, indicators, and risk factors. 

The INTERNIST-1, QMR, and PATHFINDER systems are among the first examples of BN disease 

models; these systems associated with noisy-OR networks had just illness and symptom nodes, or 

BN2Onetworks. Figure 2 (a) shows that three-layer BNs have recently supplanted these two-layer 

models. In addition to including illness risk indicators, these models reduce reliance assumptions. The 

models for which the results will be derived are easily generalisable to others, regardless of the 

complexity of their relationships. 

 

Figure 2 (a) Three-layer BN Method 

In causal inference, Structural Causal Models (SCMs), sometimes called Functional Causal Models or 

Structural Equation Models, supersede BNs. A lot of research and practical experience with SCMs has 

gone into understanding their relationship to other methods like BNs and probabilistic graphical 

models. 

An essential feature of SCM is that they depict all variables as probabilistic combinations of their 

immediate causes and an unobserved external noise component that stands in for all non-model causes. 

There is a probability distribution across the observed variables due to the fact that the state of the 

noise component is unknown.  



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 32 No. 3 (2025) 

 

571 https://internationalpubls.com 

3.2 Dual-OR Noisy Diagnostic Networks 

Building disease models sometimes involves making extra assumptions about the model beyond what 

is suggested by the DAG architecture. Among these, noisy-OR models are the most popular. Because 

they represent common sense assumptions about the relationships between symptoms and illnesses, 

noisy-OR models find widespread use in medical modelling. Furthermore, they provide effective 

reasoning and learning, and they enable a multitude of characteristics that scale linearly with the 

network size to characterise massive BNs. 

A parent Di may activate their kid S using the noisy-OR assumption if two conditions are met: (i) the 

parent is turned on, (ii) 𝑫𝒊 = 1, and (iii) the activation procedure does not fail at random. There is no 

correlation between any of the other model parameters and the failure probability, which is sometimes 

written as 𝝀𝑫𝒊, 𝑺. 

The noisy-OR component activates the kid if either parent successfully activates it. The value of S is 

the Boolean OR operation of its parents' functions of activation, 𝒔 = 𝒗𝒊𝒇(𝒅𝒊, 𝒖𝒊), where 𝒇(𝒅𝒊, 𝒖𝒊) =

𝒅𝒊˄𝒖𝒊̅̅̅, ˄, denotes the Boolean AND function,  𝒅𝒊𝝐{𝟎, 𝟏} is the state of a parent, and 𝒖𝒊𝝐{𝟎, 𝟏}  is a 

latent noise variable with a chance of failure 𝑷(𝒖𝒊 = 𝟏) = 𝝀𝑫𝒊, 𝑺. If a symptom can be activated with 

a single activation and activating a disease does not activate a symptom, then the noisy-OR model 

intuitively reflects this circumstance.  

In order to calculate the predicted sufficiency and disablement for these models, they use the twin-

networks approach for calculating counterfactuals. For the purpose of computing counterfactual 

probabilities using traditional inference methods, this approach integrates real and counterfactual 

parameters into a single SCM twin network. Abduction remains unsolvable for large SCMs, but this 

approach drastically reduces the inferential cost of computing counterfactuals.  

4. Results 

The studies comparing posterior inference with the predicted disablement and sufficiency are detailed 

here, using the models described above. Presenting the test set, which consists of a group of physicians 

and a collection of clinical scenarios. These algorithms are then tested on a variety of diagnostic tasks 

for evaluation. 

4.1 Databases and Diagnostic Method 

The usage of EHRs is one method for verifying diagnostic algorithms. When diagnostic mistakes lead 

to incorrectly labelled data, a major drawback of this method is how hard it is to define the ground 

truth diagnosis. This problem is especially noticeable in differential diagnoses because of the high rates 

of diagnostic ambiguity and miscommunication, the wide variety of applicant diseases and their 

associated diagnosis labels, the lack of or incorrect documentation of case details, and biases like the 

diagnosing doctor's expertise and education. To address these concerns, it is common practice to 

examine clinical vignettes or simulated diagnostic situations when evaluating doctors.  

A clinical vignette comprises a non-exhaustive compilation of evidence, including indicators, medical 

records, and essential demographic information such as age and birth gender, illustrating a typical 

patient's manifestation of a problem. This approach has been effective in evaluating physicians as 

individuals and juxtaposing their precision with symptoms checker technologies; it is also more 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 32 No. 3 (2025) 

 

572 https://internationalpubls.com 

resistant to inaccuracies and biases than actual data sets such as electronic health records (EHRs). 

Practising a disease depending on its established characteristics is more straightforward than doing a 

differential diagnosis. 

A different group of physicians, all with medical training at least up to the general practitioner level, 

created 1672 clinical vignettes for us to utilise as a test. Every effort has been made to ensure that the 

symptoms and risk variables align with the statistical illness model. However, to keep the study 

unbiased, the scenes include any extra clinical details that the clinicians in the studies may have as 

case notes. Each case narrative is written by one doctor and then double-checked by other physicians 

to make sure it depicts a real-life diagnostic scenario. The algorithm conceals the actual condition in 

each vignette and then uses the information from the vignettes to rank all of the modelled diseases and 

provide a diagnosis. The associative method uses the posterior to rank diseases, whereas the 

counterfactual algorithms use the predicted disablement or expected sufficiency. A partly sorted list of 

potential illnesses is the independent differential diagnosis that doctors provide. 

To guarantee that any modification in diagnostic correctness is related to the rank query used, the 

associative and counterfactual algorithms use the same disease models in all investigations. The BN is 

defined by a coalition of medical professionals and epidemiologists. Epidemiological data are used to 

determine the prior likelihood of diseases and risk variables, while several independent medical 

specialists are interviewed to determine the conditional probabilities.  

4.2 Comparison of Associative and Counterfactual Rankings 

In the first experiment, we evaluate the diagnostic accuracy of rating illnesses by posterior anticipated 

disablement and expected sufficiency. The top-k correctness is defined as the ratio of the 1672 

diagnosis vignettes in which the actual disease is included in the top-k rankings. 

. For each of the 1672 vignettes, the top-k ranked diagnoses are computed, with k = 1..., 20. The 

findings are shown in Figure 3. The accuracy for all k on the test set is about the same while using the 

anticipated disablement and expected sufficiency, hence they only provide the consequences for the 

expected adequacy for clarity's sake.  

 

Figure 3 Top k Ranked Accuracy Findings. 

The counterfactual method outperforms the associative algorithm by 2.6% when k = 1, giving the top-

ranking illness. Both methods function differently for k > 1, with the counterfactual strategy 

significantly outperforming the associative algorithm in terms of error rate reduction. With k > 5, the 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 32 No. 3 (2025) 

 

573 https://internationalpubls.com 

counterfactual approach outperforms the associative algorithm in reducing misdiagnoses by around 

31%. This indicates that although the posterior does a decent job of finding the best candidate illness, 

the counterfactual ranking does a much better job of identifying the next most probable diseases. When 

making a differential diagnosis, triage, and deciding on the best course of therapy, these secondary 

candidate disorders take on added significance. 

Finding the genuine disease's place in the rankings is a quick and easy way to compare two rankings. 

Among all 1672 vignettes, we discovered that in 24.8% of cases, the real sickness was rated higher by 

the counterfactual algorithm compared to the associative method, and in 2% of cases, it was ranked 

lower. Compared to the associative method's 3.82 plus or minus 5.26, the genuine illness is rated at 

3.17 plus or minus 4.5 using the counterfactual approach, which is a significant improvement. 

The vignettes are organised according to the genuine disease's past incidence rates: extremely frequent, 

common, uncommon, rare, and very rare. For both common and uncommon illnesses, the 

counterfactual algorithm outperforms the associative method. However, for very uncommon diseases, 

the enhancement is much more pronounced, with the counterfactual approach attaining a superior 

ranking for 33 per cent of these vignettes and rare disorders for 29.3 per cent. Since uncommon 

diseases are notoriously difficult to detect, this development is crucial since they include a wide range 

of dangerous illnesses for which diagnostic mistakes may have devastating effects. 

4.3 Compared to Doctors 

In the following study, they evaluate the associative and counterfactual algorithms with a group of 45 

medical professionals. Every doctor gets a set of 51 vignettes to work with. They then have to 

independently diagnose each vignette using a partly ranked list containing k illnesses, wherein the 

dimensions of the list k are determined by the doctor for each instance. Given a doctor and a set of 

vignettes that need a diagnosis, the associated and counterfactual algorithms are fed the same data and 

each one gives back a top-k evaluation, where k is the doctor's size. To evaluate the algorithms' and 

physicians' accuracy, they match their precision for each vignette. This way, researchers don't have to 

force clinicians to diagnose a specific number of illnesses. This is significant because, as the diagnostic 

vignette displays the physicians' level of doubt, the magnitude k of their diagnosis will naturally 

fluctuate.  

The full outcomes for all 45 doctors as well as the predicted sufficiency, posterior, and expected 

disablement grading processes. The associated and counterfactual algorithms are compared to each 

doctor's accuracy in Figure 4. For each of the 45 doctors, one can see their average accuracy; this is 

based on the percentage of vignettes in which they correctly identified the underlying illness. The 

equivalent algorithm's accuracy in diagnosing the identical vignette and returning differences of the 

same magnitude as that doctor is displayed against this measure. If both the doctors and the algorithms 

are quite accurate, then the case sets with the simplest vignettes will have the best results. On the other 

side, when it comes to more difficult vignettes that both the clinicians and the computer struggle with, 

the algorithm usually ends up with better accuracy. This indicates that diagnostic algorithms should be 

considered a supplement to human physicians, with algorithms demonstrating superior performance 

in scenarios where human error is more prevalent. 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 32 No. 3 (2025) 

 

574 https://internationalpubls.com 

The associative procedure achieves a mean accuracy over all trials of 72.53 + or - 2.98%, which is 

comparable to the average doctor's 71.41 plus or minus 3.02%. The algorithm outperforms 22 out of 

the doctors, ties with 1 doctor, and underperforms 22 out of the doctors. In comparison to both the 

typical doctor and the associated algorithm, the counterfactual algorithm obtains an impressive average 

accuracy of 77.26 plus or minus 2.80%. This puts it in the top 26% of physicians in the cohort.  

Contrary to fact method outperforms 33 out of the physicians, has a draw percentage of 1, and is less 

accurate than 11. 

  

Figure 4 Associated and Counterfactual Algorithms Are Compared to Each Doctor's Accuracy 

All things considered; the diagnostic accuracy is much better with the counterfactual approach 

compared to the associative algorithm. For uncommon illnesses, we find the improvement to be quite 

striking. In contrast to the counterfactual algorithm's top-quartile performance, the associative 

algorithm's results are on par with the typical doctors. 

5. Discussion with Conclusion 

Healthcare systems throughout the world face a formidable obstacle in the form of inadequate access 

to basic care and the problem of incorrect differential diagnoses. To use machine learning to the 

advantage and conquer these obstacles, they must first comprehend the diagnostic process and define 

precisely what they want the algorithms to produce. Diagnostics and associative inference have been 

confused in previous methods. The former seeks to identify the root of a patient's symptoms, while the 

latter learns associations between patient records and illness incidences, to ascertain the most probable 

diseases in the patient's community. In straightforward causal situations with a single illness, this 

method may work, but it severely limits the accuracy of these procedures. When a doctor must pick 

between competing disease theories. To get over these limitations, researchers need to reevaluate 

diagnostic criteria and algorithm design from the ground up. 

In this paper, researchers define diagnosis causally differently and contend that diagnosis is essentially 

an exercise in counterfactual inference. Anticipated disablement and anticipated sufficiency are two 

counterfactual diagnostic measures that have been devised. To calculate these measures, they have 

introduced a new class of diagnostic models called twin diagnostic networks. Utilising pre-existing 

diagnostic models, researchers proved that, in comparison to conventional associative rankings, using 

these counterfactual metrics to rank illness hypotheses substantially enhances diagnosis accuracy. 

When it came to expert clinical accuracy, the counterfactual procedure was in the top 26 per cent of 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 32 No. 3 (2025) 

 

575 https://internationalpubls.com 

the sample, whereas the associative algorithm was around average. In 29.3% of instances, the 

counterfactual procedure ranked the genuine illness higher than the associated process; in 33% of these 

situations, the improvement is especially noticeable for very uncommon diseases, where diagnostic 

mistakes are usually more numerous and more severe. The best part is that the disease model doesn't 

even need to be changed to take advantage of this enhancement. Existing Bayesian diagnostic methods, 

whether in or out of the medical field, may benefit from the algorithm's instant improvement thanks to 

its backward compatibility. 

In contrast to previous efforts that have centred on enhancing model architectures or tapping into new 

data sources, findings provide a novel way to improve clinical decision systems at the expert level by 

altering model queries to take causal information into account. These results provide credence to the 

claim that, in certain fields, human specialists will always be able to outperform machine learning 

approaches that do not use causal reasoning. The current tests have concentrated on comparing the 

algorithms to doctors, but future trials might find out how well algorithms work as medical support 

systems that help doctors by giving them a second opinion when diagnosing patients. The combined 

diagnosis of a doctor and the method will probably be more reliable than either one alone, as the system 

seems to complement human physicians by doing better in scenarios that humans have trouble 

diagnosing. 

Reference: 

[1] Richens, Jonathan G., Ciarán M. Lee, and Saurabh Johri. "Counterfactual diagnosis." arXiv preprint 

arXiv:1910.06772 (2019). 

[2] Richens, Jonathan G., Ciarán M. Lee, and Saurabh Johri. "Improving the accuracy of medical diagnosis with causal 

machine learning." Nature Communications 11.1 (2020): 3923. 

[3] Prosperi, Mattia, et al. "Causal inference and counterfactual prediction in machine learning for actionable healthcare." 

Nature Machine Intelligence 2.7 (2020): 369-375. 

[4] Verma, Sahil, et al. "Counterfactual explanations and algorithmic recourses for machine learning: A review." ACM 

Computing Surveys (2020). 

[5] Singla, Sumedha, et al. "Explaining the black box smoothly—a counterfactual approach." Medical image analysis 84 

(2023): 102721. 

[6] Lin, Junfan, et al. "Towards causality-aware inferring: a sequential discriminative approach for medical diagnosis." 

IEEE Transactions on Pattern Analysis and Machine Intelligence (2023). 

[7] Wu, Hang, et al. "Clinical decision making under uncertainty: a bootstrapped counterfactual inference approach." 

BMC Medical Informatics and Decision Making 24.1 (2024): 1-15. 

[8] Thiagarajan, Jayaraman J., et al. "Training calibration-based counterfactual explainers for deep learning models in 

medical image analysis." Scientific Reports 12.1 (2022): 597. 

[9] AlJalaud, Ebtisam, and Manar Hosny. "Counterfactual Explanation of AI Models using an Adaptive Genetic 

Algorithm with Embedded Feature Weights." IEEE Access (2024). 

[10] Temraz, Mohammed, and Mark T. Keane. "Solving the class imbalance problem using a counterfactual method for 

data augmentation." Machine Learning with Applications 9 (2022): 100375. 

[11] Verma, Sahil, John Dickerson, and Keegan Hines. "Counterfactual explanations for machine learning: A review." 

arXiv preprint arXiv:2010.10596 2 (2020): 1. 

[12] Kelly, Luke, et al. "Evolving Visual Counterfactual Medical Imagery Explanations with Cooperative Co-evolution 

using Dynamic Decomposition." Proceedings of the Genetic and Evolutionary Computation Conference Companion. 

2024. 



Communications on Applied Nonlinear Analysis 

ISSN: 1074-133X 

Vol 32 No. 3 (2025) 

 

576 https://internationalpubls.com 

[13] Atad, Matan, et al. "Counterfactual Explanations for Medical Image Classification and Regression using Diffusion 

Autoencoder." arXiv preprint arXiv:2408.01571 (2024). 

[14] Dai, Xinyue, et al. "Counterfactual explanations for prediction and diagnosis in XAI." Proceedings of the 2022 

AAAI/ACM Conference on AI, Ethics, and Society. 2022. 

[15] Cheng, Furui, Yao Ming, and Huamin Qu. "Dece: Decision explorer with counterfactual explanations for machine 

learning models." IEEE Transactions on Visualization and Computer Graphics 27.2 (2020): 1438-1447. 

[16] Huang, Hao, Emetis Niazmand, and Maria-Esther Vidal. "Hybrid AI Approach for Counterfactual Prediction over 

Knowledge Graphs for Personal Healthcare." Artificial Intelligence and Data Science for Healthcare: Bridging Data-

Centric AI and People-Centric Healthcare. 2024. 

[17] Chou, Yu-Liang, et al. "Counterfactuals and causability in explainable artificial intelligence: Theory, algorithms, and 

applications." Information Fusion 81 (2022): 59-83. 

[18] Mertes, Silvan, et al. "Ganterfactual—counterfactual explanations for medical non-experts using generative 

adversarial learning." Frontiers in artificial intelligence 5 (2022): 825565. 

[19] Singh, Chaitanya, et al. "Applied machine tool data condition to predictive smart maintenance by using artificial 

intelligence." International Conference on Emerging Technologies in Computer Engineering. Cham: Springer 

International Publishing, 2022.. 

[20] Dash, Saloni, Vineeth N. Balasubramanian, and Amit Sharma. "Evaluating and mitigating bias in image classifiers: 

A causal perspective using counterfactuals." Proceedings of the IEEE/CVF Winter Conference on Applications of 

Computer Vision. 2022. 

[21] Son, Jaemin, et al. "An interpretable and interactive deep learning algorithm for a clinically applicable retinal fundus 

diagnosis system by modelling finding-disease relationship." Scientific Reports 13.1 (2023): 5934. 

[22] Barraza, Joaquín Figueroa, Enrique López Droguett, and Marcelo Ramos Martins. "FS-SCF network: Neural network 

interpretability based on counterfactual generation and feature selection for fault diagnosis." Expert Systems with 

Applications 237 (2024): 121670. 

[23] Zhang, Baoliang, et al. "Counterfactual inference graph network for disease prediction." Knowledge-Based Systems 

255 (2022): 109722. 

[24] Sun, Zhaohong, et al. "Adversarial reinforcement learning for dynamic treatment regimes." Journal of Biomedical 

Informatics 137 (2023): 104244. 

[25] Patro, R. Azhagumurugan, R. Sathya, K. Kumar, T. R. Kumar and M. V. S. Babu, "A hybrid approach estimates the 

real-time health state of a bearing by accelerated degradation tests, Machine learning," 2021 Second International 

Conference on Smart Technologies in Computing, Electrical and Electronics (ICSTCEE), Bengaluru, India, 2021, pp. 

1-9, doi: 10.1109/ICSTCEE54422.2021.9708591