Evaluation of Orthogonal Vector Projection Method in ST Algorithm

for Generating Differential Diagnoses of Chest Pain: A Pilot Study

Irosh Fernando

, Luke Nepia

, Hoang Mai Khanh Do

and Edward Holmes

Hunter New England Area Health, 72 Watt St, Newcastle, NSW, Australia

Faculty of Medicine and Health, University of New England, Elm Avenue, Armidale NSW 2351, Australia

School of Medicine and Public Health, Faculty of Health and Medicine, University of Newcastle, Australia

Keywords: Medical Expert Systems, Computer Algorithms for Medical Diagnosis, Select and Test (ST) Algorithm,

Orthogonal Vector Projection Method, Algorithm for Diagnosing Chest Pain.

Abstract: Diagnosing chest pain can be a challenging process with potential misdiagnoses causing significant morbidity

and mortality, and the associated healthcare cost and burden. As a potential solution to increase the diagnostic

accuracy and rule out non-life-threatening conditions, we have evaluated the method known as orthogonal

vector projection which is a part of the Select and Test (ST) algorithm for medical diagnosis, as a pilot study.

Using a knowledgebase consisting of 12 diagnoses and 43 clinical features, we have evaluated 47 clinical

cases by comparing the diagnosis given by a senior clinician to the diagnosis arrived by the orthogonal vector

projection method.

1 INTRODUCTION

Chest pain is common physical complaint with a

lifetime prevalence of 25% in the general population,

resulting in common presentations to emergency

departments (Thomsett et al, 2018), (Chew et al,

2016), (Cullen et al, 2015). While there are many

possible causes of chest pain ranging from benign

causes to life-threatening medical emergencies such

as acute coronary syndrome (ACS), discriminating

them can be difficult (Cullen et al, 2015), (Geyser et

al, 2016). For example, it is known that between 50%

to 80% of the time, patients with possible ACS are

misdiagnosed and sent home without proper

treatment (Geyser eta l, 2016); and about one third of

patients who didn’t have a diagnosis related to their

chest pain, are known to be later diagnosed with ACS

or die from cardiovascular disease (Fordyce et al,

2016). On the other hand, the benign causes of chest

pain still often require evaluation including

investigations amounting to healthcare cost of

diagnosis which burdens patients and health care

services (Cullen et al, 2015). Furthermore, clinicians

are known to make diagnostic errors due to number

of factors including fatigue and time pressure. Hence,

use of diagnostic algorithm to improve diagnostic

https://orcid.org/0000-0003-1239-9277

accuracy, mitigate the errors, and minimise

unnecessary investigations, is highly desirable.

In this research work, we have used the method

known as orthogonal vector projection of ST

algorithm, which was introduced by (Fernando et al,

2016) and has been evaluated in generating

differential diagnoses for psychiatric conditions.

In this study, two different evaluations were

done to explore the potential use of the method for

triaging (i.e., arriving at differential diagnoses prior

to conducting investigations) and a diagnostic tool

(i.e., arriving at diagnosis with all clinical features

including investigation results).

2 ORTHOGONAL VECTOR

PROJECTION METHOD

(OVPM)

A given clinical presentation with a set of clinical

features, requiring a diagnosis, is conceptualised as a

binary vector in which, each feature is assigned a

binary value to indicate if the feature is present or not

in the patient. On the other hand, each potential

diagnosis presented as a real vector corresponding to

Fernando, I., Nepia, L., Do, H. and Holmes, E.

Evaluation of Orthogonal Vector Projection Method in ST Algorithm for Generating Differential Diagnoses of Chest Pain: A Pilot Study.

DOI: 10.5220/0012456000003657

In Proceedings of the 17th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2024) - Volume 2, pages 695-699

ISBN: 978-989-758-688-0; ISSN: 2184-4305

695

the set of clinical features, with each real element

corresponding to respective clinical feature, being a

relative weight assigned according to its diagnostic

importance (i.e., a clinical feature which is highly

indicative of the diagnosis is given a higher weight

compared to one which is less indicative of the

diagnosis). Using these vectors, a similarity measure

with respect to each diagnosis is derived as follows.

Let 𝑛 be the total number of clinical features,

𝐿



⃗

⎝

⎜

⎛

𝑙



𝑙



⎠

⎟

⎞

where,each element 𝑙



∈



0,1



be the vector of clinical features,

𝐷



⃗

⎝

⎜

⎛

𝑑



𝑑



⎠

⎟

⎞

where,each element 𝑑



∈[0,1],

be a potential diagnosis (i.e., a column vector in the

knowledge base) which satisfies,

𝑑







=1,

and we derive, 𝑋

⃗

the Hadamard product (element-

wise product)

𝑋

⃗

=𝐿



⃗

⃘𝐷



⃗

Then using the similarity measure denoted as 𝑟 is

derived as follows,

𝑟=

𝑋

⃗

.𝐷



⃗

𝐷



⃗





Where 𝐷



⃗



is the length of the vector 𝐷



⃗

Suppose 𝑚 is the number of total diagnoses, then the

knowledgebase is a 𝑛×𝑚 matrix which consists of

the column vectors corresponding to each diagnosis.

3 STUDY DESIGN AND DATA

COLLECTION

A total of 12 diagnoses were carefully chosen along

with the list of 43 clinical features consisting of

physical symptoms, clinical examination findings and

investigations. With the view of two potential use of

the method as a triage tool, a separate set of vectors

for diagnoses were created excluding investigation

findings and redistributing the weights. The weights

were assigned to each clinical feature for each

diagnosis subjectively using clinical expertise and

adjusted using test cases. The two knowledgebases

are tabulated in the appendix.

Assuming a prevalence between 5-10% of the

chest pain related diagnoses in emergency department,

and a predetermined sensitivity and specificity of 80%

for each and individual diagnosis, the estimated

sample size was 980 participants, which was not

feasible to achieve with the time and resources

available. However, combining all diagnoses as one

general diagnostic entity with their aggregated

prevalence be more than 95%, with predetermined

sensitivity and specificity of 92% for the general

diagnostic entity, and with 8% of maximum marginal

error and 95% confidence level, the required sample

size was deemed as only 47 participants.

Therefore, evaluation of the orthogonal vector

projection method was conducted as a pilot study to

determine the general sensitivity and specificity for

combined diagnoses as opposed to determining

sensitivity and specificity for each individual

diagnosis.

The diagnostic data (i.e., list of clinical features

and the diagnosis given by a senior clinician) from the

patients who were recruited for the study, was

collected after obtaining ethics approval from Hunter

New England Local Area Health District (John

Hunter Hospital and Maitland Hospital) where the

study was conducted over a period of 4 months. The

recruitment data are summarised in Table-1.

Table 1: Chest pain related diagnoses and number of

clinical cases.

Diagnosis Number of cases

STEMI 6

NSTEMI 23

Unstable angina 4

Pulmonary embolism 4

Pneumonia 9

Gastric ulcer 1

Aortic dissection 0

Pericarditis 0

Pneumothorax 0

Cholecystitis 0

Costochondritis 0

Panic attack 0

HEALTHINF 2024 - 17th International Conference on Health Informatics

696

4 ANALYSIS AND RESULTS

For each clinical case, the similarity measure 𝑟 was

calculated for each for each 12 diagnoses. We chose

arbitrarily the following set of cut-off points for

positive diagnoses: 0.6𝑟,0.7𝑟,0.8𝑟. The

analysis was conducted separately for the diagnoses

with investigations and without investigations.

Without investigation results being included, the

OVPM was able to achieve specificity above 90% at

all cut-off points with a negative predictive value of

90%, and the maximum sensitivity achieved was

70%. On the other hand, with the investigations being

included, the OVPM was able to achieve 87.2%

(95% CI 74.3%-95.2%) sensitivity and 99.2% (95%

CI 98.0%-99.8%) specificity for the cut-off 0.6>r;

and the positive predictive value and the negative

predictive values were 93% and 97.5% respectively;

Table 2 summarises the results for the three most

common diagnoses in the data set.

Table 2: Sensitivity and specificity for the cut-off 0.6>r

with investigation results being included.

STEMI

STEMI Pneumonia

True + 6 21 8

False + 0 1 1

False - 0 2 1

True - 43 23 37

Sensitivit

100% 91.3% 88.9%

Specificit

100% 95.8% 97.4%

The area under the receiver operating

characteristic curve (AUROC) for the results of the

analysis without and with investigation results were

0.772 and 0.928 respectively (Figures 1 and 2).

Figure 1: Receiver operating characteristic curve for

investigation results being excluded.

Figure 2: Receiver operating characteristic curve for

investigation results being included.

5 DISCUSSION

This pilot study shows that the OVPM has a great

potential in triaging chest pain and diagnosis; the

results have shown to have an excellent diagnostic

accuracy as per the expected standards (Šimundić et

al, 2009). Particularly, with high specificity and NPV,

OVPM has the potential to use as a triage tool with

the utility of ruling out certain diagnoses and thus

minimising the cost of unnecessary investigations.

There were number of limitations in the study.

Firstly, whilst deriving the optimum weights for each

clinical feature for each diagnosis is critical for

accurate results, it was done subjectively as a manual

process; thus, not necessarily representing the

optimum weights. Secondly, the sample size of the

study was small and couldn’t recruit patients for all

the diagnoses, having no patients for 6 diagnoses out

of the 12 diagnoses chosen.

Future areas of research involve developing an

automated process of deriving weights for each pair

of clinical feature and diagnosis; also conducting

further evaluation with bigger sample size.

REFERENCES

Thomsett R, Cullen L. The assessment and management of

chest pain in primary care: A focus on acute coronary

syndrome. Aust J Gen Pract.2018;47(5):246-251.

Chew DP, Scott IA, Cullen L, French JK, Briffa TG,

Tideman PA, et al. National Heart Foundation of

Australia and Cardiac Society of Australia and New

Zealand: Australian clinical guidelines for the

Evaluation of Orthogonal Vector Projection Method in ST Algorithm for Generating Differential Diagnoses of Chest Pain: A Pilot Study

697

management of acute coronary syndromes 2016. Med J

Aust. 2016 Aug 1;205(3):128-33.

Cullen L, Greenslade J, Merollini K, Graves N, Hammett

CJ, Hawkins T, et al. Cost and outcomes of assessing

patients with chest pain in an Australian emergency

department. Med J Aust. 2015 May;202(8):427-32.

Geyser M, Smith S. Chest pain prevalence, causes, and

disposition in the emergency department of a regional

hospital in Pretoria. Afr J Prim Health Care Fam Med.

2016;8(1):e1-5.

Fordyce CB, Newby DE, Douglas PS. Diagnostic Strategies

for the Evaluation of Chest Pain: Clinical Implications

From SCOT-HEART and PROMISE. J Am Coll

Cardiol. 2016;67(7):843-852.

Fernando IP, Henskens FA. The select and test algorithm

for inference in medical diagnostic reasoning:

Implementation and evaluation in clinical psychiatry.

2016 IEEE/ACIS 15

International Conference on

Computer and Information Science (ICIS).

2016;1(1):1-9.

Šimundić AM. Measures of Diagnostic Accuracy: Basic

Definitions. EJIFCC. 2009 Jan 20;19(4):203-11.

PMID: 27683318; PMCID: PMC4975285.

APPENDIX

Figure 3: Knowledgebase consisting of clinical features and their assigned values for each possible differential diagnosis,

excluding investigations; unstable angina (UA), aortic dissection (AD), pericarditis (PC), pulmonary embolism (PE),

pneumothorax (PT), pneumonia (Pneum), cholecystitis (Chole), peptic ulcer disease (PUD).

HEALTHINF 2024 - 17th International Conference on Health Informatics

698

Figure 4: Knowledgebase consisting of clinical features and their assigned values for each possible differential diagnosis,

including investigations.

Evaluation of Orthogonal Vector Projection Method in ST Algorithm for Generating Differential Diagnoses of Chest Pain: A Pilot Study

699