Applying Machine Learning Techniques to Identify Undiagnosed Patients with Exocrine Pancreatic Insufficiency

BACKGROUND

Exocrine pancreatic insufficiency (EPI) is a serious condition characterized by a lack of functional exocrine pancreatic enzymes and the resultant inability to properly digest nutrients. EPI can be caused by a variety of disorders, including chronic pancreatitis, pancreatic cancer, and celiac disease. EPI remains underdiagnosed because of the nonspecific nature of clinical symptoms, lack of an ideal diagnostic test, and the inability to easily identify affected patients using administrative claims data.

OBJECTIVES

To develop a machine learning model that identifies patients in a commercial medical claims database who likely have EPI but are undiagnosed.

METHODS

A machine learning algorithm was developed in Scikit-learn, a Python module. The study population, selected from the 2014 Truven MarketScan® Commercial Claims Database, consisted of patients with EPI-prone conditions. Patients were labeled with 290 condition category flags and split into actual positive EPI cases, actual negative EPI cases, and unlabeled cases. The study population was then randomly divided into a training subset and a testing subset. The training subset was used to determine the performance metrics of 27 models and to select the highest performing model, and the testing subset was used to evaluate performance of the best machine learning model.

RESULTS

The study population consisted of 2088 actual positive EPI cases, 1077 actual negative EPI cases, and 437 530 unlabeled cases. In the best performing model, the precision, recall, and accuracy were 0.91, 0.80, and 0.86, respectively. The best-performing model estimated that the number of patients likely to have EPI was about 12 times the number of patients directly identified as EPI-positive through a claims analysis in the study population. The most important features in assigning EPI probability were the presence or absence of diagnosis codes related to pancreatic and digestive conditions.

CONCLUSIONS

Machine learning techniques demonstrated high predictive power in identifying patients with EPI and could facilitate an enhanced understanding of its etiology and help to identify patients for possible diagnosis and treatment.

Earthquake nucleation by transient deformations caused by the M = 7.9 Denali, Alaska, earthquake

The permanent and dynamic (transient) stress changes inferred to trigger earthquakes are usually orders of magnitude smaller than the stresses relaxed by the earthquakes themselves, implying that triggering occurs on critically stressed faults. Triggered seismicity rate increases may therefore be most likely to occur in areas where loading rates are highest and elevated pore pressures, perhaps facilitated by high-temperature fluids, reduce frictional stresses and promote failure. Here we show that the 2002 magnitude M = 7.9 Denali, Alaska, earthquake triggered widespread seismicity rate increases throughout British Columbia and into the western United States. Dynamic triggering by seismic waves should be enhanced in directions where rupture directivity focuses radiated energy, and we verify this using seismic and new high-sample GPS recordings of the Denali mainshock. These observations are comparable in scale only to the triggering caused by the 1992 M = 7.4 Landers, California, earthquake, and demonstrate that Landers triggering did not reflect some peculiarity of the region or the earthquake. However, the rate increases triggered by the Denali earthquake occurred in areas not obviously tectonically active, implying that even in areas of low ambient stressing rates, faults may still be critically stressed and that dynamic triggering may be ubiquitous and unpredictable.

Earthquake triggering by seismic waves following the Landers and Hector Mine earthquakes

The proximity and similarity of the 1992, magnitude 7.3 Landers and 1999, magnitude 7.1 Hector Mine earthquakes in California permit testing of earthquake triggering hypotheses not previously possible. The Hector Mine earthquake confirmed inferences that transient, oscillatory 'dynamic' deformations radiated as seismic waves can trigger seismicity rate increases, as proposed for the Landers earthquake. Here we quantify the spatial and temporal patterns of the seismicity rate changes. The seismicity rate increase was to the north for the Landers earthquake and primarily to the south for the Hector Mine earthquake. We suggest that rupture directivity results in elevated dynamic deformations north and south of the Landers and Hector Mine faults, respectively, as evident in the asymmetry of the recorded seismic velocity fields. Both dynamic and static stress changes seem important for triggering in the near field with dynamic stress changes dominating at greater distances. Peak seismic velocities recorded for each earthquake suggest the existence of, and place bounds on, dynamic triggering thresholds. These thresholds vary from a few tenths to a few MPa in most places, depend on local conditions, and exceed inferred static thresholds by more than an order of magnitude. At some sites, the onset of triggering was delayed until after the dynamic deformations subsided. Physical mechanisms consistent with all these observations may be similar to those that give rise to liquefaction or cyclic fatigue.

Triggering of earthquake aftershocks by dynamic stresses

It is thought that small 'static' stress changes due to permanent fault displacement can alter the likelihood of, or trigger, earthquakes on nearby faults. Many studies of triggering in the near-field, particularly of aftershocks, rely on these static changes as the triggering agent and consider them only in terms of equivalent changes in the applied load on the fault. Here we report a comparison of the aftershock pattern of the moment magnitude Mw = 7.3 Landers earthquake, not only with static stress changes but also with transient, oscillatory stress changes transmitted as seismic waves (that is, 'dynamic' stresses). Dynamic stresses do not permanently change the applied load and thus can trigger earthquakes only by altering the mechanical state or properties of the fault zone. These dynamically weakened faults may fail after the seismic waves have passed by, and might even cause earthquakes that would not otherwise have occurred. We find similar asymmetries in the aftershock and dynamic stress patterns, the latter being due to rupture propagation, whereas the static stress changes lack this asymmetry. Previous studies have shown that dynamic stresses can promote failure at remote distances, but here we show that they can also do so nearby.

Seismicity Remotely Triggered by the Magnitude 7.3 Landers, California, Earthquake

The magnitude 7.3 Landers earthquake of 28 June 1992 triggered a remarkably sudden and widespread increase in earthquake activity across much of the western United States. The triggered earthquakes, which occurred at distances up to 1250 kilometers (17 source dimensions) from the Landers mainshock, were confined to areas of persistent seismicity and strike-slip to normal faulting. Many of the triggered areas also are sites of geothermal and recent volcanic activity. Static stress changes calculated for elastic models of the earthquake appear to be too small to have caused the triggering. The most promising explanations involve nonlinear interactions between large dynamic strains accompanying seismic waves from the mainshock and crustal fluids (perhaps including crustal magma).

Surgical revascularization of left main coronary artery stenosis: determinants of perioperative and long-term outcome in the 1980s

The postoperative courses of 176 patients who underwent coronary artery bypass surgery for significant left main coronary artery stenosis were analyzed to determine which preoperative clinical and angiographic factors correlated best with outcome. Clinical variables included age, sex, New York Heart Association (NYHA) anginal class, presence of unstable angina, and surgical class. The angiographic variables included percentage of left main stenosis, presence of right coronary artery stenosis, coronary dominance, number of vessels diseased, myocardial jeopardy score, and ejection fraction. The overall perioperative mortality rate was 9.1%. There was a significant increase in perioperative mortality among female patients (p less than 0.05) and patients undergoing emergency surgery (p less than 0.05). Patients with left main stenosis of 80% or more or with balanced or left dominant circulation showed trends toward increased perioperative mortality. Life-table analysis showed that emergency surgery and left main stenosis of 80% or more correlated with increased long-term mortality (p less than 0.05). No other variable tested showed a significant correlation with either perioperative or long-term mortality. A comparison of these results with studies performed in the 1970s shows that there has been considerable change in those factors which place a patient at increased risk for mortality during surgical treatment of left main coronary artery stenosis.

Simultaneous contrast imaging of the left ventricle by two-dimensional echocardiography and standard ventriculography

Simultaneous contrast two-dimensional echocardiograms and contrast ventriculograms were analyzed for 19 cardiac cycles in 6 patients. Ventriculographic volume was underestimated by 10 +/- 4.5% (p less than .001) by the contrast echocardiograms, despite good correlation (r = 0.88). Discrepancies could not be assigned to changes in volume between studies nor to a lack of precise endocardial definition. Simultaneous imaging demonstrated that the ultrasonic transducer was located 33 degrees cephalad to the cardiac apex. Although angiographic volume was severely underestimated, ejection fraction (r = 0.93) and stroke volume (r = 0.90) could be calculated from two-dimensional echocardiograms using regression equations.