Better Living Through Better Prediction

It’s Minority Report again – but pre-health, instead of pre-crime.

This is work from Sinai, where they are applying unsupervised machine learning techniques to generate intelligence from the electronic health record.  In less technical terms, they’re building SkyNet – for good.

These authors selected patients with at least five records in their data warehouse in the years leading up to 2013 as their modeling cohort.  Based on these ~700,000 patients, they abstracted all encounters, diagnoses, clinical notes, medications, and structured order data.  These various data types were then pared down to approximately 41,000 features that neither appeared in >80% of the records nor in fewer than five records, and then these features were normalized for analysis.  The novelty in their approach was their specific unsupervised data abstraction, reducing each patient to a dense vector of 500 features.  They then selected patients with at least one new ICD-9 diagnosis recored in their EHR in 2014, and divided them as their validation and test cohorts for disease prediction.

The results varied by diagnosis, but, most importantly, demonstrated their method appears superior to several other methods of abstraction – a non-abstracted “raw features” analysis, principal component analysis, Gaussian mixture model, k-means, and independent component analysis.  Using a random forest model for prediction, their abstraction method – “DeepPatient” – provided the best substrate for future diagnoses.  For example, their method worked best on “diabetes mellitus with complications”, providing an AUC for this diagnosis of 0.907.  Other high-scoring disease predictions including various cancers, cardiovascular disorders, and mental health issues.

Much work remains to be completed before similar technology is applicable in a practical clinical context.  This application does not even specifically account for the actual value of lab tests, only prediction of outcomes based on the co-occurence of other clinical features with a lab test result present.  Prediction strength also varied greatly by disease process; it is likely a more restricted or lightly supervised model will outperform their generic unsupervised general model with regard to specific short-term outcomes relating to emergency care.  And, of course, even when such models are being developed, they will still require testing and practice refinement regarding the traditional challenges balancing accuracy, risk tolerance, and resource utilization.

“Deep Patient: An Unsupervised Representation to Predict the Future of Patients from the Electronic Health Records”

Severe Sepsis … or ß-Agonist

As our sepsis overlords entrenched new “quality measures” and other protocol-driven resuscitation requirements in our Emergency Departments, this article serves as a lovely reminder of the importance of staying cognitively engaged.

Lactate levels can be elevated by metabolic and microcirculatory derangements related to the spectrum of sepsis – but also other, non-infectious causes.  These include hepatic disease, multiple toxodromes, and multiple medications – one of the most commonly used being beta-agonist therapy for obstructive airways.  This very simple study examines the physiologic changes in healthy volunteers receiving 10mg of nebulized albuterol, as compared with nebulized saline.  Placebo volunteers had no change in lactate or placebo.  Albuterol receiving volunteers had an average increase in lactate of 0.77 mmol/L and an average decrease in potassium of 0.5 mEq/L.  Lactate increases, however, were highly variable – ranging from 0.04 to 2.02 mmol/L.

These data aren’t perfectly generalizable to the critically or pseudo-critically ill, but they’re a reasonable starting point for a gross estimate.  They’re also justification for reconsideration of potentially inappropriate therapies for an intermediate-range lactate that obstinately refuses to clear – in the context of receiving multiple rounds of nebulizers.

At the very least, it’s a reminder of the various exceptions to our protocols we need to consider to prevent costly and avoidable harms.

“The Effect of Nebulized Albuterol on Serum Lactate and Potassium in Healthy Subjects”
https://www.ncbi.nlm.nih.gov/pubmed/26857949

You’re Drugged By What You Eat

When you drink sewage, it’s pretty clear what you’re getting: sewage, with all the expect, overt nutritional implications.

When you eat fruit, you expect a tasty and healthy meal.

Unfortunately, as this proof-of-concept study reveals, when you eat fruit watered by sewage, you get a little from Column A and a little from Column B.

This study conducted in Israel evaluated the presence of carbamazapine in healthy university student volunteers.  As carbamazapine is recognized to be excreted as an unchanged active drug, it is ubiquitous in untreated sewage – read “wastewater”.  Half the recruited students were then randomized to eat a selection of vegetables acquired from producers watering their fields with wastewater, rather than treated or fresh water.  Unsurprisingly, the various vegetables and leafy greens from the wastewater farm had elevated levels of carbamazepine – and so did the participants randomized to that group.

Just a happy thought for the day – and a reason to think ever so slightly more about from where your food’s water comes.

“Human Exposure to Wastewater-Derived Pharmaceuticals in Fresh Produce: A Randomized Controlled Trial Focusing on Carbamazepine”
http://pubs.acs.org/doi/abs/10.1021/acs.est.5b06256

Put the Platelets Away in ICH

Sometimes, medical practice in the setting of uncertainty simply turns out to be futile and low-value.

This is one of those times where we’ve probably been at least futile, and possibly harmful.

Life-threatening or critical intracranial bleeding in the setting of concomitant antiplatelet therapy frequently offers a dire prognosis.  As part of our standard “don’t just stand there!” approach in Emergency Medicine, patients with ICH in this setting are frequently transfused platelets in an effort to provide untainted clotting substrate.  This practice, however, has never been reinforced by substantiated evidence, and the pharmacokinetics of the antiplatelet agents suggests this strategy is unlikely to be efficacious.

This is the PATCH trial, a randomized, open-label trial conducted at 60 hospitals between 2009 and 2015, investigating the utility of platelet transfusion in the setting of ICH.  Patients with normal baseline functional status and ICH while taking aspirin, clopidogrel, or dipyridamole were eligible for inclusion.  Specific excluded ICH were epidural or subdural hematomas, significant intraventricular blood, surgical intervention planned, or those in which death appeared imminent.  Treating clinicians could not be blinded to study arm allocation, but follow-up assessors and data analysis was masked.  The primary outcome is was functional outcome on the modified Rankin Scale, analyzed via ordinal shift analysis.

The authors do not present the number of patients screened for potential enrollment during the study period, but, ultimately, 190 participants were included from 41 centers.  The authors state patients were well-balanced on most demographics, although median ICH volumes were a little higher in the platelet-transfusion group, with 34% of patients having ICH >30mL versus only 21% in the standard-care group.  There were four patients in the platelet-transfusion group who did not receive transfusion, and two in the standard-care that did.

In the end, outcomes were universally dismal.  Only 15 patients in the entire study survived with minimal disability or better.  The vast majority of patients were at least moderately disabled or dead at follow-up.  And, while the confidence intervals for many of their comparisons cross unity, none of the trends favored platelet transfusion.  Generally speaking, there were more deaths, fewer patients with minimal disability, and additional adverse events in the transfusion group.

I tend to feel this is a small enough cohort the heterogeneity between individual patients is enough to effect the overall results – including the apparent harms relating to platelet-transfusion.  However, there is certainly no signal of benefit, and lacking a compelling indication to utilize a scarce resource, I believe this is enough to suggest this practice should be routinely avoided.

“Platelet transfusion versus standard care after acute stroke due to spontaneous cerebral haemorrhage associated with antiplatelet therapy (PATCH): a randomised, open-label, phase 3 trial”
http://www.thelancet.com/journals/lancet/article/PIIS0140-6736(16)30392-0/abstract

Triaging Large Artery Occlusions

Endovascular intervention for acute stroke can be quite useful – in appropriately selected patients.  However, few centers are capable of such interventions, and the technology to properly angiographically evaluated for large-artery occlusions is not available in all settings.  Thus, it is just as critical for patients to be clinically screened in some fashion to prevent over-utilization of scarce resources.

These authors retrospectively reviewed 1,004 acute stroke patients admitted to their facility since 2008, 328 of which had large-vessel occlusions: ICA, M1, or basilar artery.  They calculated the accuracy, sensitivity, and specificity of multiple different potential clinical scoring systems, cut-offs.  Unfortunately, every score made some trade-off – either in the rate of false-negative results excluding patients from potential intervention, or in the rate of false-positive results serving to simply subject every patient to advanced imaging.  The maximum accuracy of all their various scores topped around around 78%.

The authors’ conclusions are reasonable, if a little limited.  They feel every patient presenting with an acute stroke within 6 hours of symptom onset should undergo vascular imaging.  These are both reasonable, but ignore one of the major uses for simple clinical scoring systems: prehospital triage.  Admitting none of these are perfect, _something_ must be put to use – and, probably, given the current bandwidth for endovascular intervention, something with the highest specificity.

For what it’s worth, we use RACE to triage for CT perfusion, but CPSSS, ROSIER, or just NIHSS cut-offs around 10 would all be fair choices.

“Clinical Scales Do Not Reliably Identify Acute Ischemic Stroke Patients With Large-Artery Occlusion”
https://www.ncbi.nlm.nih.gov/pubmed/27125526

Tying Procalcitonin to Critical Care

It has been hard, over the years, to truly identify a role for procalcitonin.  Generally speaking, its best niche seems to be as a sort of C-reactive protein on steroids – a non-specific infectious or inflammatory marker with better sensitivity than WBC.  This has led to some usage in zero-miss contexts such as neonatal sepsis, as well as a potential role in antibiotic stewardship.

These authors, many of which are supported by the manufacturers of the procalcitonin assay, evaluate its predictive power in the setting of pneumonia hospitalization, attempting to risk-stratify patients for the combined endpoint of vasopressor support or invasive ventilation.  Their goal, they say, is to use procalcitonin levels to better inform level-of-care decisions – both escalated and de-escalated – at the time of hospital admission.

They analyzed 1,770 patients from a prior pneumonia study for whom banked serum samples were adequate for procalcitonin measurement, 115 of whom met their combined critical illness endpoint.  They report risk of critical illness increased approximately linearly with procalcitonin from 4% when procalcitonin was undetectable, to 22.4% when procalcitonin was 10ng/mL or above.  The AUC for procalcitonin alone was 0.69, as compared to WBC at 0.54.  Then, they further go on to add usage of procalcitonin in conjunction with other risk-stratification scores – ATS minor criteria, PSI, and SMART-COP – provided additional discriminatory information.

This could be a potentially useful and interesting application of procalcitonin – except they don’t really make any comparisons to other available tools, other than a straw man comparison with WBC.  Would the venerable CRP have a similar AUC?  Or, better yet, a lab we already use nearly ubiquitously to detect occult severe sepsis – a lactic acid level?  The authors do not present any specific discussion of alternative approaches – of which their friends at BioMerieux probably appreciate.

“Procalcitonin as an Early Marker of the Need for Invasive Respiratory or Vasopressor Support in Adults with Community-Acquired Pneumonia”
https://www.ncbi.nlm.nih.gov/pubmed/27107491

Prime Time for Dexamethasone in Asthma?

We adore dexamethasone here at EMLoN headquarters.  A pharmacy stocked solely with ketamine, droperidol, and dexamethasone could carry you far in life.  Unfortunately, because of our established bias, the challenge then arises to reconcile the actual results of a trial with what is so desperately wished to be true.

This is a randomized, double-blinded, non-inferiority trial comparing single-dose dexamethasone with five days of prednisone in the treatment of asthma with acute exacerbation.  Single-dose and double-dose dexamethasone protocols have been evaluated in pediatric trials of reasonable size, but data in adults has been lacking.  However, many clinicians – including myself – have assumed generalizability of the pediatric findings to adults, and have been using single-dose dexamethasone protocols for years.

If only one takeaway can be had from this trial, it is: never, ever, skimp on sample size. Far easier said than done, of course, but due to the complex structure and assumptions required for non-inferiority trials, this is a negative trial.  There were 465 subjects randomized – but 89 excluded from analysis as either subsequent admissions or as lost to follow-up.  Of the remaining 376 patients, 9.8% of the prednisone group demonstrated Emergency Department recidivism compared with 12.1% of dexamethasone.  This 2.3% difference between groups, however, suffers a 95% of -4.1 to 8.6%, and exceeds the pre-determined clinically-relevant non-inferiority margin of 8%.

So, unfortunately, we still do not have a precise estimate for the effectiveness difference between a prednisone-based strategy and dexamethasone.  Considering the healthcare burden of asthma in our Emergency Departments, it is somewhat surprising we still have such a paucity of data – as even a small difference in effectiveness may have profound effects on Emergency Department utilization.  In the end, it comes down to where you lay on the spectrum of pre-study odds for non-inferiority or equivalence.  For me, the guarantee of compliance with treatment derived from a single-dose of dexamethasone outweighs the continued uncertainty over its true effectiveness, and this study gives me no cause to change my practice.

“A Randomized Controlled Noninferiority Trial of Single Dose of Oral Dexamethasone Versus 5 Days of Oral Prednisone in Acute Adult Asthma”
https://www.ncbi.nlm.nih.gov/pubmed/27117874

Endovascular for Stroke – Even Better than the Evidence

What happens when you let Medtronic, et al, author an article on endovascular therapy in The Lancet:  exactly what you’d expect.

We are, in principle, fans of endovascular therapy for acute stroke as presented in the major trials: ESCAPE, EXTEND-IA, and SWIFT-PRIME.  These trials carefully selected eligible patients by use of advanced perfusion imaging and demonstrated high rates of revascularization.  Viable brain plus restored flow has face validity for improved outcomes.

However, these sponsored authors use the meta-analysis for its most nefarious purpose: to obfuscate the important subtleties and eligibility criteria of its included trials.  These authors pool the aforementioned trials, along with MR CLEAN and REVASCAT to provide the following conclusion:

“Endovascular thrombectomy is of benefit to most patients with acute ischaemic stroke caused by occlusion of the proximal anterior circulation, irrespective of patient characteristics or geographical location.”(emphasis mine)

The authors also provide a staggering number-needed-to-treat for endovascular therapy of 2.6.

But, of course, this was written to shock and awe the lay press and general medicine community, rather than edify the astute clinician.  Their NNT is not based on the typical dichotomous cut-off used in stroke trials of mRS 0-1 or 0-2 – but rather the hopelessly flawed ordinal shift analysis.  As the decades turn, apparently, we have forgotten why this approach was frowned upon from the start: it is not appropriate to equate the outcome value difference between mRS 5 and 4 with the difference between mRS 3 and 2, and the limitations in inter-rater reliability in the mRS introduce a vast additional amount of measurement error.  Then, by burying any mention of the strict imaging criteria responsible for the bulk of benefit seen in these trials, they mislead the reader into considering this therapy appropriate for all-comers.

Is there any value to these data as presented?  A little.  There is hypothesis generating evidence that tPA prior to endovascular therapy provides no additive benefit.  There is also evidence that increasingly distal sites of occlusion may not benefit from intervention.

Unfortunately, the flaws in this article outweigh the few potentially usable insights.  This is just yet another piece of direct-to-physician marketing masquerading as scientific evidence.

“Endovascular thrombectomy after large-vessel ischaemic stroke: a meta-analysis of individual patient data from five randomised trials”
https://www.ncbi.nlm.nih.gov/pubmed/26898852

How Many ED Visits are Truly Inappropriate?

I’ve seen quite a bit of feedback on social media regarding this research letter in JAMA Internal Medicine.

This study evaluated, using National Hospital Ambulatory Medical Care Survey data, the incidence of hospital admission stratified by triage Emergency Severity Index.  They analyzed 59,293 representative visits from the sample and found 7.5% of them, on a weighted basis, were categorized as “non-urgent” – an ESI level 5 or presumed equivalent.  The typical assumption regarding these non-urgent visits is they represent inappropriate Emergency Department utilization.  This study found, however:

“… a nontrivial proportion of ED visits that were deemed nonurgent arrived by ambulance, received diagnostic services, had procedures performed, and were admitted to the hospital, including to critical care units.”

There are always limitations regarding the NHAMCS data, particularly with missing and imputed data.  Based on this, I tend to feel these data lack face validity.  The weighted incidence of admission for non-urgent patients was 4.4% compared with 12.8% of urgent visits, while 0.7% of non-urgent visits were to critical care units compared with 1.3% of urgent visits.  I certainly do not see similar relative proportions of admission, and then to critical care, for level 5 patients in my multiple practice environments.

Regardless, the general implication made by these authors is probably reasonable, refuting usage of ESI triage level 5 to accurately represent inappropriate Emergency Department visits.  However, left equally unstated, is an acknowledgement that ESI also fails to accurately categorize urgent visits – which ties to the rhetoric of trying to conflate “non-urgent” as “inappropriate and “urgent” as “appropriate”.

ESI, as currently implemented, will not be a reliable tool for directing patients to other sources of care – but, with some fuzziness, probably still gives a reasonable estimate of the overall burden of inappropriate ED visits for some policy applications.

“Urgent Care Needs Among Nonurgent Visits to the Emergency Department”
https://www.ncbi.nlm.nih.gov/pubmed/27089549