AI and clinical diagnosis

see also:

• artificial intelligence

• AI is likely to have important roles in supporting clinical diagnosis

• humans, like all animals, are primed to rapidly recognise patterns ONCE they have learned those patterns - and this requires experience / exposure and awareness
• most conditions have patterns that can be recognised by varying degrees of medical expertise (or by AI training on those patterns)
• some clinical features are pathognomonic of a specific condition (ie. have 100% specificity for that condition)

• most lay people can recognise common conditions such as:
  • deformed broken long bone
  • thermal burn
  • sharp object causing cuts or stab wounds
  • bruises from blunt injury
  • acute allergic reactions
  • viral upper respiratory illness
  • atopic dermatitis
  • acne vulgaris
  • hay fever
  • “gastroenteritis”
  • “constipation”

• pathognomonic clinical features of fractures
• cellulitis
• abscess
• common presentations of many common conditions

• tension pneumothorax
• acute pulmonary oedema
• welding flash burns to eyes
• appendicitis
• renal colic
• biliary colic
• basic ECG patterns
• common presentations of most common conditions

• ascending cholangitis
• aortic dissection
• “Christmas eye”
• complex ECG patterns
• uncommon presentations of common conditions

• hidden features on Xrays, MRIs, ECGs, photos, genomic analysis, etc

• this requires learning or perhaps online searching of differential diagnoses for the presenting clinical feature
• it also requires extracting from the patient the correct assessment of the clinical feature(s)
• this is often the first part of the clinician's assessment after taking a history and examination and once formulated a range of methods can be utilised as outlined below to narrow down the most likely diagnosis
  • re-assessment of clinical features specifically looking for less obvious features of each differential
  • investigations as per algorithmic approach, shotgun approach, or a Bayesian approach

• order a range of investigations and see what they provide
• eg. extensive range of blood tests such as FBE, U&E, LFTs, lipase, glucose, CRP, D-Dimer, troponin, HCG as well as ECG, urine test, CXR
• this can be extremely useful when these are easily available, cost effective and particularly if there are difficulties in obtaining adequate clinical assessment such as language issues, unconscious patient, or the clinical picture is confusing
• this approach has the advantage of throwing up unexpected but life threatening results such as a CRP of 250 with cold sepsis
• the downside is that the more investigations you have at hand for no specific indication, the more likely you will find “red herring” abnormal results which send you down time wasting and costly rabbit holes if you don't have the experience to be able to ignore them eg. raised D-Dimer when there is a likely non-thrombotic cause
• AI could potentially be of use if it can be trained on all this investigation data as well as clinical presentation data, age, gender, etc PLUS actual diagnoses and clinical outcomes for very large number of clinical presentations as it is potentially excellent at determining likely outcomes from patterns which may not be self-evident to humans

• this is often employed in emergent time poor situations where efficient, relatively consistent diagnostic approaches relatively independent of level of experience can be used to rule out most immediate life threatening conditions or to provide a cost-effective method to approach presentations which may have a range of differential diagnoses
• it still generally requires integration of a range of data such as:
  • presenting features
  • risk factors - past history, family history, occupation, lifestyle and cultural factors, genomics (if available, otherwise ethnicity factors), etc
  • potential precipitant events
• often though, a patient will get to the end of the algorithm and no cause will have been found which often creates much angst and either this status is accepted and the condition is allowed to play out - either resolve or get worse, or further diagnostic processes need to be considered such as specialist referral
• examples:
  • chest pain
    • check the ECG for clues - is it a STEMI? is there evidence of PE? is there evidence of pericarditis?
    • could it be a tension pneumothorax - clinical features
    • is it really primarily chest pain or is it coming from the abdomen?
    • could it be aortic dissection - pain description, risk factors, BP each arm, CXR, consider aortogram if clinical features suggestive
    • could it be a PE - PERC rule if no recent Covid or other confounding events, consider D-Dimer or bedside echo, consider CTPA
    • could it be a NSTEMI - serial troponins, etc
    • is it really chest wall pain - rash of shingles, clinical rib fracture, etc
• currently AI is poor at this approach as LLMs deal primarily with word associations and NOT with algorithms, nor cause and effect

• this is often subconsciously or consciously utilised in algorithmic and pattern recognition approaches
• it requires an understanding of:
  • ability to generate a differential diagnosis and create a pre-test probability of each
  • prevalence of each condition in a differential diagnosis for that patient population
  • sensitivity and specificity of any planned investigation for a specific condition
• an example of how this is best used is whether or not to do a CTPA for suspected PE
  • if your pre-test probability for PE is very low, even a positive CTPA is likely to be a false positive rather than a true positive - this concept is difficult for many to understand
• currently AI is poor at this approach as LLMs deal primarily with word associations and currently do not understand disease prevalence, nor sensitivity and specificity of investigations