Using natural language processing for automated classification of disease and to identify misclassified ICD codes in cardiac disease

Aims

ICD codes are used for classification of hospitalizations. The codes are used for administrative, financial, and research purposes. It is known, however, that errors occur. Natural language processing (NLP) offers promising solutions for optimizing the process. To investigate methods for automatic classification of disease in unstructured medical records using NLP and to compare these to conventional ICD coding.

Methods and results

Two datasets were used: the open-source Medical Information Mart for Intensive Care (MIMIC)-III dataset (n = 55.177) and a dataset from a hospital in Belgium (n = 12.706). Automated searches using NLP algorithms were performed for the diagnoses 'atrial fibrillation (AF)' and 'heart failure (HF)'. Four methods were used: rule-based search, logistic regression, term frequency-inverse document frequency (TF-IDF), Extreme Gradient Boosting (XGBoost), and Bio-Bidirectional Encoder Representations from Transformers (BioBERT). All algorithms were developed on the MIMIC-III dataset. The best performing algorithm was then deployed on the Belgian dataset. After preprocessing a total of 1438 reports was retained in the Belgian dataset. XGBoost on TF-IDF matrix resulted in an accuracy of 0.94 and 0.92 for AF and HF, respectively. There were 211 mismatches between algorithm and ICD codes. One hundred and three were due to a difference in data availability or differing definitions. In the remaining 108 mismatches, 70% were due to incorrect labelling by the algorithm and 30% were due to erroneous ICD coding (2% of total hospitalizations).

Conclusion

A newly developed NLP algorithm attained a high accuracy for classifying disease in medical records. XGBoost outperformed the deep learning technique BioBERT. NLP algorithms could be used to identify ICD-coding errors and optimize and support the ICD-coding process.

Deaths caused by osteoporotic fractures in Japan: An epidemiological study

BACKGROUND

Osteoporosis is a global issue with a worldwide prevalence of 18.3%, and the presence of coexisting fragility fractures can reduce the survival rate by approximately 20%. In Japan, the prevalence of osteoporosis is estimated to be 12.8 million, and the annual occurrence of hip fractures is approximately 193,400. Remarkably, coexisting hip or spinal fragility fractures caused by slight external force meet the Japanese diagnostic criterion for osteoporosis regardless of bone mineral density. However, only 191 deaths due to osteoporosis were published in 2021 in Japan. With the concern that some cases of hip and spinal fragility fractures were assigned an underlying cause of death of traumatic fracture instead of osteoporosis, this study aimed to elucidate the actual number of deaths due to osteoporosis in Japan.

METHODS

We used the data from Japan in 2018. First, the number of deaths due to osteoporosis and hip or spinal fractures was reviewed using published vital statistics. Second, we calculated the number of elderly deaths (age ≥80 years) resulting from hip or spinal fractures caused by falls on the same level using data from approximately 1.4 million annual individual death certificates. Combining the above data, the actual number of deaths due to osteoporosis was estimated.

RESULTS

Only 190 deaths due to osteoporosis were reported in the published data. The individual certificate data revealed 3437 elderly deaths due to hip or spinal fractures caused by falls on the same level, which could meet the criteria of osteoporotic fragility fractures. Accordingly, the estimated number of deaths caused by osteoporosis was calculated as 3,627, approximately 19 times the published value.

CONCLUSIONS

After researching the individual death certificate data focusing on the coexisting hip or spinal fragility fracture, it was implied that osteoporosis may have a higher mortality rate in Japan than what is published.

Analysis of causes of death among brought-in-dead cases in a third-level Hospital in Lusaka, Republic of Zambia, using the tariff method 2.0 for verbal autopsy: a cross-sectional study

BACKGROUND

Over one third of deaths in Zambian health facilities involve someone who has already died before arrival (i.e., Brough in Dead), and in most BiD cases, the CoD have not been fully analyzed. Therefore, this study was designed to evaluate the function of automated VA based on the Tariff Method 2.0 to identify the CoD among the BiD cases and the usefulness by comparing the data on the death notification form.

METHODS

The target site was one third-level hospital in the Republic of Zambia's capital city. All BiD cases who reached the target health facility from January to August 2017 were included. The deceased's closest relatives were interviewed using a structured VA questionnaire and the data were analyzed using the SmartVA to determine the CoD at the individual and population level. The CoD were compared with description on the death notification forms by using t-test and Cohen's kappa coefficient.

RESULTS

One thousand three hundred seventy-eight and 209 cases were included for persons aged 13 years and older (Adult) and those aged 1 month to 13 years old (Child), respectively. The top CoD for Adults were infectious diseases followed by non-communicable diseases and that for Child were infectious diseases, followed by accidents. The proportion of cases with a determined CoD was significantly higher when using the SmartVA (75% for Adult and 67% for Child) than the death notification form (61%). A proportion (42.7% for Adult and 46% for Child) of the CoD-determined cases matched in both sources, with a low concordance rate for Adult (kappa coefficient = 0.1385) and a good for Child(kappa coefficient = 0.635).

CONCLUSIONS

The CoD of the BiD cases were successfully analyzed using the SmartVA for the first time in Zambia. While there many erroneous descriptions on the death notification form, the SmartVA could determine the CoD among more BiD cases. Since the information on the death notification form is reflected in the national vital statistics, more accurate and complete CoD data are required. In order to strengthen the death registration system with accurate CoD, it will be useful to embed the SmartVA in Zambia's health information system.

Correlating injury severity scores and major trauma volume using a state-wide in-patient administrative dataset linked to trauma registry data-A retrospective analysis from New South Wales Australia

INTRODUCTION

Trauma registries are used to analyse and report activity and benchmark quality of care at designated facilities within a trauma system. These capabilities may be enhanced with the incorporation of administrative and electronic medical record datasets, but are currently limited by the use of different injury coding systems between trauma and administrative datasets.

OBJECTIVES

Use an Abbreviated Injury Scale to International Classification of Disease (AIS-ICD) mapping tool to correlate estimated injury severity scores and major trauma volume based on administrative data collections with trauma registry data.

METHODS

Adult trauma cases were identified from the New South Wales Trauma Registry between 2012 and 2016 and linked probabilistically using age, facility and date of facility arrival to the Admitted Patient Data Collection (APDC). Estimated Injury Severity Scores (ISS) were derived using the AIS-ICD mapping tool applied to diagnoses contained in the APDC.

RESULTS

A total of eligible 13,439 cases were analysed. The overall correlation between trauma registry ISS and ISS estimated from APDC using the AIS-ICD mapping tool was low to moderate (Spearman Rho 0.41 95%CI 0.40, 0.43). Based on an estimated ISS cut-off value of 8, there was high correlation between estimated trauma volume and the number of major trauma cases at each facility (Spearman Rho 0.98, 95%CI 0.95, 0.99). Trauma Revised Injury Severity Score (TRISS) was associated with only slightly higher mortality prediction performance compared to estimated ISS (AUROC 0.76 95%CI 0.75, 0.78 versus AUROC 0.74 95%CI 0.73, 0.76).

CONCLUSION

A low to moderate correlation exists between individual patient ISS scores based on AIS to ICD mapping of in-patient data collection, but a high correlation for overall major trauma volume using the AIS-ICD mapping at facility level with comparable TRISS mortality prediction.

Accuracy of an administrative database for pancreatic cancer by international classification of disease 10th codes: A retrospective large-cohort study

BACKGROUND

Korean National Health Insurance (NHI) claims database provides large-cohort. However, studies regarding accuracy of administrative database for pancreatic cancer (PC) have not been reported. We aimed to identify accuracy of NHI database regarding PC classified by international classification of disease (ICD)-10 codes.

AIM

To identify the accuracy and usefulness of administrative database in PC and the accurate ICD codes for PC with location.

METHODS

Study and control groups were collected from 2003 to 2016 at Seoul National University Bundang Hospital. Cases of PC were identified in NHI database by international classification of diseases, 10th revision edition (ICD-10 codes) supported with V codes. V code is issued by medical doctors for covering 95% of medical cost by Korean government. According to pathologic reports, definite or possible diagnoses were defined using medical records, images, and pathology.

RESULTS

A total of 1846 cases with PC and controls were collected. Among PC, only 410 (22.2%) cases were identified as specific cancer sites including head in 234 (12.7%) cases, tail in 104 (5.6%) cases and body in 72 (3.9%) cases. Among PC, 910 (49.3%) cases were diagnosed by definite criteria. Most of these were adenocarcinoma (98.0%). The rates of definite diagnosis of PC were highest in head (70.1%) followed by body (47.2%) and tail (43.3%). False-positive cases were pancreatic cystic neoplasm and metastasis to the pancreas. In terms of the overall diagnosis of PC, sensitivity, specificity, positive predictive value, and negative predictive value were 99.95%, 98.72%, 98.70%, and 99.95%, respectively. Diagnostic accuracy was similar both in terms of diagnostic criteria and tumor locations.

CONCLUSION

Korean NHI claims database collected according to ICD-10 code with V code for PC showed good accuracy.

Calcium channel blockers as drug repurposing candidates for gestational diabetes: Mining large scale genomic and electronic health records data to repurpose medications

New therapeutic approaches are needed for gestational diabetes mellitus (GDM), but must show safety and efficacy in a historically understudied population. We studied associations between electronic medical record (EMR) phenotypes and genetic variants to uncover drugs currently considered safe in pregnancy that could treat or prevent GDM. We identified 129 systemically active drugs considered safe in pregnancy targeting the proteins produced from 196 genes. We tested for associations between GDM and/or type 2 diabetes (DM2) and 306 SNPs in 130 genes represented on the Illumina Infinium Human Exome Bead Chip (DM2 was included due to shared pathophysiological features with GDM). In parallel, we tested the association between drugs and glucose tolerance during pregnancy as measured by the glucose recorded during a routine 50-g glucose tolerance test (GTT). We found an association between GDM/DM2 and the genes targeted by 11 drug classes. In the EMR analysis, 6 drug classes were associated with changes in GTT. Two classes were identified in both analyses. L-type calcium channel blocking antihypertensives (CCBs), were associated with a 3.18 mg/dL (95% CI -6.18 to -0.18) decrease in glucose during GTT, and serotonin receptor type 3 (5HT-3) antagonist antinausea medications were associated with a 3.54 mg/dL (95% CI 1.86-5.23) increase in glucose during GTT. CCBs were identified as a class of drugs considered safe in pregnancy could have efficacy in treating or preventing GDM. 5HT-3 antagonists may be associated with worse glucose tolerance.

Exploration of association rule mining for coding consistency and completeness assessment in inpatient administrative health data

OBJECTIVE

Data quality assessment is a challenging facet for research using coded administrative health data. Current assessment approaches are time and resource intensive. We explored whether association rule mining (ARM) can be used to develop rules for assessing data quality.

MATERIALS AND METHODS

We extracted 2013 and 2014 records from the hospital discharge abstract database (DAD) for patients between the ages of 55 and 65 from five acute care hospitals in Alberta, Canada. The ARM was conducted using the 2013 DAD to extract rules with support ≥0.0019 and confidence ≥0.5 using the bootstrap technique, and tested in the 2014 DAD. The rules were compared against the method of coding frequency and assessed for their ability to detect error introduced by two kinds of data manipulation: random permutation and random deletion.

RESULTS

The association rules generally had clear clinical meanings. Comparing 2014 data to 2013 data (both original), there were 3 rules with a confidence difference >0.1, while coding frequency difference of codes in the right hand of rules was less than 0.004. After random permutation of 50% of codes in the 2014 data, average rule confidence dropped from 0.72 to 0.27 while coding frequency remained unchanged. Rule confidence decreased with the increase of coding deletion, as expected. Rule confidence was more sensitive to code deletion compared to coding frequency, with slope of change ranging from 1.7 to 184.9 with a median of 9.1.

CONCLUSION

The ARM is a promising technique to assess data quality. It offers a systematic way to derive coding association rules hidden in data, and potentially provides a sensitive and efficient method of assessing data quality compared to standard methods.

Validation of an International Classification of Disease, Ninth Revision coding algorithm to identify decompressive craniectomy for stroke

Background

Although International Classification of Disease, Ninth Revision, Clinical Modification (ICD9-CM) coding is the basis of administrative claims data, no study has validated an ICD9-CM algorithm to identify patients undergoing decompressive craniectomy for space-occupying supratentorial infarction.

Methods

Patients who underwent decompressive craniectomy for stroke at our institution were retrospectively identified and their associated ICD9-CM codes were extracted from billing data. An ICD9-CM algorithm was generated and its accuracy compared against physician review.

Results

A total of 10,925 neurosurgical operations were performed from December 2008 to March 2015, of which 46 (0.4%) were decompressive craniectomy for space-occupying stroke. The ICD9-CM procedure code for craniectomy (01.25) was only encoded in 67.4% of patients, while craniotomy (01.24) was used in 19.6% and lobectomy (01.39, 01.53, 01.59) in 13.1%. The ICD-9-CM algorithm included patients with a diagnosis codes for cerebral infarction (433.11, 434.01, 434.11, and 434.91) and a procedure code for craniotomy, craniectomy, or lobectomy. Patients were excluded with an ICD9-CM diagnosis code for brain tumor, intracranial abscess, subarachnoid hemorrhage, vertebrobasilar infarction, intracranial aneurysm, Moyamoya disease, intracranial venous sinus thrombosis, vertebral artery dissection, congenital cerebrovascular anomaly, head trauma or an ICD9-CM procedure code for laminectomy. This algorithm had a sensitivity of 97.8%, specificity of 99.9%, positive predictive value of 88.2%, and negative predictive value of 99.9%. The majority of false-positive results were patients who underwent evacuation of a primary intracerebral hematoma.

Conclusion

An ICD-9-CM algorithm based on diagnosis and procedure codes can effectively identify patients undergoing decompressive craniectomy for supratentorial stroke.

Design and implementation of a national public health surveillance system in Jordan

Understanding and improving the health status of communities depend on effective public health surveillance. Adoption of new technologies, standardised case definitions and clinical guidelines for accurate diagnosis, and access to timely and reliable data, remains a challenge for public health surveillance systems however and existing public health surveillance systems are often fragmented, disease specific, inconsistent and of poor quality. We describe the application of an enterprise architecture approach to the design, planning and implementation of a national public health surveillance system in Jordan. This enabled a well planned and collaboratively supported system to be built and implemented using consistent standards for data collection, management, reporting and use. The system is case-based and integrated and employs mobile information technology to aid collection of real-time, standardised data to inform and improve decision-making at different levels of the health system.

The Protective Effect of Hispanic Ethnicity on Chronic Obstructive Pulmonary Disease Mortality is Mitigated by Smoking Behavior

RATIONALE

Studies in Hispanic/Latino populations have shown ethnicity to be either a predictive or protective factor for COPD mortality and it is unclear whether this disagreement is attributable to ethnic differences in smoking rates, smoking behavior or differences in genetic susceptibility.

OBJECTIVES

This study will examine the role of smoking behavior as a means of explaining differences in risk for COPD mortality between Hispanics and non-Hispanic whites.

METHODS

Participants were recruited into a cohort study from the San Luis Valley in Colorado beginning in the early 1980's and followed for mortality until 2012. COPD and cardiovascular disease are often comorbid conditions and account for the competing risk of CVD in the analysis of COPD mortality. Mortality searches were conducted regularly and all ICD codes were collected for mortality event. Primary and secondary causes of each event were assessed using appropriate codes.

RESULTS

Hispanic current smokers did not differ from NHW current smokers in years smoked (p=0.6) but Hispanic former smokers accumulated more years smoked compared to NHW former smokers (22 vs. 20, p=0.047). Hispanic ethnicity was significantly protective for COPD mortality adjusting for age, gender, pre-existing emphysema, hypertension and smoking status and accounting for the effect of CVD mortality (RR=0.58, 95% C.I. 0.34-0.99, p=0.035). Further adjustment for smoking behavior mitigated this effect.

CONCLUSIONS

The lower COPD mortality seen in Hispanic smokers may be due lower cumulative exposure to tobacco smoke. Thus, smoking behavior may play a key role in explaining differences in COPD mortality as they relate to Hispanic ethnicity.