Applying machine learning to identify pediatric patients with newly diagnosed acute lymphoblastic leukemia using administrative data.

Case identification in administrative databases is challenging as diagnosis codes alone are not adequate for case ascertainment. We utilized machine learning (ML) to efficiently identify pediatric patients with newly diagnosed acute lymphoblastic leukemia. We tested nine ML models and validated the best model internally and externally. The optimal model had 97% positive predictive value (PPV) and 99% sensitivity in internal validation; 94% PPV and 82% sensitivity in external validation. Our ML model identified a large cohort of 21,044 patients, demonstrating an efficient approach for cohort assembly and enhancing the usability of administrative data.

Association of the social disorganization index with time to first septic shock event in children with acute myeloid leukemia.

BACKGROUND: Pediatric acute myeloid leukemia (AML) chemotherapy increases the risk of life-threatening complications, including septic shock (SS). An area-based measure of social determinants of health, the social disorganization index (SDI), was hypothesized to be associated with SS and SS-associated death (SS-death). METHODS: Children treated for de novo AML on two Children's Oncology Group trials at institutions contributing to the Pediatric Health Information System (PHIS) database were included. The SDI was calculated via residential zip code data from the US Census Bureau. SS was identified via PHIS resource utilization codes. SS-death was defined as death within 2 weeks of an antecedent SS event. Patients were followed from 7 days after the start of chemotherapy until the first of end of front-line therapy, death, relapse, or removal from study. Multivariable-adjusted Cox regressions estimated hazard ratios (HRs) comparing time to first SS by SDI group. RESULTS: The assembled cohort included 700 patients, with 207 (29.6%) sustaining at least one SS event. There were 233 (33%) in the SDI-5 group (highest disorganization). Adjusted time to incident SS did not statistically significantly differ by SDI (reference, SDI-1; SDI-2: HR, 0.84 [95% confidence interval (CI), 0.51-1.41]; SDI-3: HR, 0.70 [95% CI, 0.42-1.16]; SDI-4: HR, 0.97 [95% CI, 0.61-1.53]; SDI-5: HR, 0.72 [95% CI, 0.45-1.14]). Nine patients (4.4%) with SS experienced SS-death; seven of these patients (78%) were in SDI-4 or SDI-5. CONCLUSIONS: In a large, nationally representative cohort of trial-enrolled pediatric patients with AML, there was no significant association between the SDI and time to SS.

Surgical neonates: A retrospective review of procedures and postoperative outcomes at a quaternary children's hospital.

INTRODUCTION: Neonates have a high incidence of respiratory and cardiac perioperative events. Disease severity and indications for surgical intervention often dovetail with an overall complex clinical course and predispose these infants to adverse long-term neurodevelopmental outcomes and increased length of stay. Our aims were to describe severe and nonsevere early postoperative complications to establish a baseline of care outcomes and to identify subgroups of surgical neonates and procedures for future prospective studies. METHODS: Electronic health record data were examined retrospectively for a cohort of patients who had general anesthesia from January 26, 2015 to August 31, 2018. Inclusion criteria were full-term infants with postmenstrual age less than 44 weeks or premature infants less than 60 weeks postmenstrual age undergoing nonimaging, noncardiac surgery. Severe postoperative complications were defined as mortality, reintubation, positive blood culture, and surgical site infection. Nonsevere early postoperative outcomes were defined as hypoglycemia, hyperglycemia, hypothermia, hyperthermia, and readmission within 30 days. RESULTS: About 2569 procedures were performed in 1842 neonates of which 10.9% were emergency surgeries. There were 120 postoperative severe complications and 965 nonsevere postoperative outcomes. Overall, 30-day mortality was 1.8% for the first procedure performed, with higher mortality seen on subgroup analysis for patients who underwent exploratory laparotomy (10.3%) and congenital lung lesion resection (4.9%). Postoperative areas for improvement included hyperglycemia (13.9%) and hypothermia (7.9%). DISCUSSION: The mortality rate in our study was comparable to other studies of neonatal surgery despite a high rate of emergency surgery and a high prevalence of prematurity in our cohort. The early outcomes data identified areas for improvement, including prevention of postoperative glucose and temperature derangements. CONCLUSIONS: Neonates in this cohort were at risk for severe and nonsevere adverse postoperative outcomes. Future studies are suggested to improve mortality and adverse event rates.

Acute and chronic kidney injury during therapy for pediatric acute leukemia: A report from the Leukemia Electronic Abstraction of Records Network (LEARN).

Children with acute leukemia are at increased risk of kidney injury. Using electronic health record data from three centers between 2010 and 2018, this study retrospectively described acute kidney injury (AKI) and chronic kidney disease (CKD) prevalence in children with acute lymphoblastic or myeloid leukemia (ALL, AML) using Common Terminology Criteria for Adverse Events (CTCAE) and Kidney Disease Improving Global Outcomes (KDIGO) definitions. AKI during therapy was 25% (ALL) and 32% (AML) using CTCAE, versus 84% (ALL) and 74% (AML) using KDIGO. CKD prevalence was low and Grade 1/Stage 2. Further investigation is needed to optimally define kidney injury in acute leukemia.

Leveraging machine learning to identify acute myeloid leukemia patients and their chemotherapy regimens in an administrative database.

BACKGROUND: Administrative datasets are useful for identifying rare disease cohorts such as pediatric acute myeloid leukemia (AML). Previously, cohorts were assembled using labor-intensive, manual reviews of patients' longitudinal chemotherapy data. METHODS: We utilized a two-step machine learning (ML) method to (i) identify pediatric patients with newly diagnosed AML, and (ii) among the identified AML patients, their chemotherapy courses, in an administrative/billing database. Using 2558 patients previously manually reviewed, multiple ML algorithms were derived from 75% of the study sample, and the selected model was tested in the remaining hold-out sample. The selected model was also applied to assemble a new pediatric AML cohort and further assessed in an external validation, using a standalone cohort established by manual chart abstraction. RESULTS: For patient identification, the selected Support Vector Machine model yielded a sensitivity of 0.97 and a positive predictive value (PPV) of 0.97 in the hold-out test sample. For course-specific chemotherapy regimen and start date identification, the selected Random Forest model yielded overall PPV greater than or equal to 0.88 and sensitivity greater than or equal to 0.86 across all courses in the test sample. When applied to new cohort assembly, ML identified 3016 AML patients with 10,588 treatment courses. In the external validation subset, PPV was greater than or equal to 0.75 and sensitivity was greater than or equal to 0.82 for patient identification, and PPV was greater than or equal to 0.93 and sensitivity was greater than or equal to 0.94 for regimen identifications. CONCLUSION: A carefully designed ML model can accurately identify pediatric AML patients and their chemotherapy courses from administrative databases. This approach may be generalizable to other diseases and databases.

Validating Use of ICD-10 Diagnosis Codes in Identifying Physical Abuse Among Young Children.

OBJECTIVE: Evaluate the positive predictive value of International Classification of Disease, 10th Revision, Clinical Modification (ICD-10-CM) codes in identifying young children diagnosed with physical abuse. METHODS: We extracted 230 charts of children <24 months of age who had any emergency department, inpatient, or ambulatory care encounters between Oct 1, 2015 and Sept 30, 2020 coded using ICD-10-CM codes suggestive of physical abuse. Electronic health records were reviewed to determine if physical abuse was considered during the medical encounter and assess the level of diagnostic certainty for physical abuse. Positive predictive value of each ICD-10-CM code was assessed. RESULTS: Of 230 charts with ICD-10 codes concerning for physical abuse, 209 (91%) had documentation that a diagnosis of physical abuse was considered during an encounter. The majority of cases, 138 (60%), were rated as definitely or likely abuse, 36 cases (16%) were indeterminate, and 35 (15%) were likely or definitely accidental injury. Other forms of suspected maltreatment were discussed in 16 (7%) charts and 5 (2%) had no documented concerns for child maltreatment. The positive predictive values of the specific ICD-10 codes for encounters rated as definitely or likely abuse varied considerably, ranging from 0.89 (0.80-0.99) for T74.12 "Adult and child abuse, neglect, and other maltreatment, confirmed" to 0.24 (95% CI: 0.06-0.42) for Z04.72 "Encounter for examination and observation following alleged child physical abuse." CONCLUSIONS: ICD-10-CM codes identify young children who experience physical abuse, but certain codes have a higher positive predictive value than others.

Absolute lymphocyte count recovery following initial acute myelogenous leukemia therapy: Implications for adoptive cell therapy.

BACKGROUND: An adequate absolute lymphocyte count (ALC) is an essential first step in autologous chimeric antigen receptor (CAR) T-cell manufacturing. For patients with acute myelogenous leukemia (AML), the intensity of chemotherapy received may affect adequate ALC recovery required for CAR T-cell production. We sought to analyze ALC following each course of upfront therapy as one metric for CAR T-cell manufacturing feasibility in children and young adults with AML. PROCEDURE: ALC data were collected from an observational study of patients with newly diagnosed AML between the ages of 1 month and 21 years who received treatment between the years of 2006 and 2018 at one of three hospitals in the Leukemia Electronic Abstraction of Records Network (LEARN) consortium. RESULTS: Among 193 patients with sufficient ALC data for analysis, the median ALC following induction 1 was 1715 cells/µl (interquartile range: 1166-2388), with successive decreases in ALC with each subsequent course. Similarly, the proportion of patients achieving an ALC >400 cells/µl decreased following each course, ranging from 98.4% (190/193) after course 1 to 66.7% (22/33) for patients who received a fifth course of therapy. CONCLUSIONS: There is a successive decline of ALC recovery with subsequent courses of chemotherapy. Despite this decline, ALC values are likely sufficient to consider apheresis prior to the initiation of each course of upfront therapy for the majority of newly diagnosed pediatric AML patients, thereby providing a window of opportunity for T-cell collection for those patients identified at high risk of relapse or with refractory disease.

Rates of laboratory adverse events by course in paediatric leukaemia ascertained with automated electronic health record extraction: a retrospective cohort study from the Children's Oncology Group.

BACKGROUND: Adverse events are often misreported in clinical trials, leading to an incomplete understanding of toxicities. We aimed to test automated laboratory adverse event ascertainment and grading (via the ExtractEHR automated package) to assess its scalability and define adverse event rates for children with acute myeloid leukaemia and acute lymphoblastic leukaemia. METHODS: For this retrospective cohort study from the Children's Oncology Group (COG), we included patients aged 0-22 years treated for acute myeloid leukaemia or acute lymphoblastic leukaemia at Children's Healthcare of Atlanta (Atlanta, GA, USA) from Jan 1, 2010, to Nov 1, 2018, at the Children's Hospital of Philadelphia (Philadelphia, PA, USA) from Jan 1, 2011, to Dec 31, 2014, and at the Texas Children's Hospital (Houston, TX, USA) from Jan 1, 2011, to Dec 31, 2014. The ExtractEHR automated package acquired, cleaned, and graded laboratory data as per Common Terminology Criteria for Adverse Events (CTCAE) version 5 for 22 commonly evaluated grade 3-4 adverse events (fatal events were not evaluated) with numerically based CTCAE definitions. Descriptive statistics tabulated adverse event frequencies. Adverse events ascertained by ExtractEHR were compared to manually reported adverse events for patients enrolled in two COG trials (AAML1031, NCT01371981; AALL0932, NCT02883049). Analyses were restricted to protocol-defined chemotherapy courses (induction I, induction II, intensification I, intensification II, and intensification III for acute myeloid leukaemia; induction, consolidation, interim maintenance, delayed intensification, and maintenance for acute lymphoblastic leukaemia). FINDINGS: Laboratory adverse event data from 1077 patients (583 from Children's Healthcare of Atlanta, 200 from the Children's Hospital of Philadelphia, and 294 from the Texas Children's Hospital) who underwent 4611 courses (549 for acute myeloid leukaemia and 4062 for acute lymphoblastic leukaemia) were extracted, processed, and graded. Of the 166 patients with acute myeloid leukaemia, 86 (52%) were female, 80 (48%) were male, 96 (58%) were White, and 132 (80%) were non-Hispanic. Of the 911 patients with acute lymphoblastic leukaemia, 406 (45%) were female, 505 (55%) were male, 596 (65%) were White, and 641 (70%) were non-Hispanic. Patients with acute myeloid leukaemia had the most adverse events during induction I and intensification II. Hypokalaemia (one [17%] of six to 75 [48%] of 156 courses) and alanine aminotransferase (ALT) increased (13 [10%] of 134 to 27 [17%] of 156 courses) were the most prevalent non-haematological adverse events in patients with acute myeloid leukaemia, as identified by ExtractEHR. Patients with acute lymphoblastic leukaemia had the greatest number of adverse events during induction and maintenance (eight adverse events with prevalence ≥10%; induction and maintenance: anaemia, platelet count decreased, white blood cell count decreased, neutrophil count decreased, lymphocyte count decreased, ALT increased, and hypocalcaemia; induction: hypokalaemia; maintenance: aspartate aminotransferase [AST] increased and blood bilirubin increased), as identified by ExtractEHR. 187 (85%) of 220 total comparisons in 22 adverse events in four AAML1031 and six AALL0923 courses were substantially higher with ExtractEHR than COG-reported adverse event rates for adverse events with a prevalence of at least 2%. INTERPRETATION: ExtractEHR is scalable and accurately defines laboratory adverse event rates for paediatric acute leukaemia; moreover, ExtractEHR seems to detect higher rates of laboratory adverse events than those reported in COG trials. These rates can be used for comparisons between therapies and to counsel patients treated on or off trials about the risks of chemotherapy. ExtractEHR-based adverse event ascertainment can improve reporting of laboratory adverse events in clinical trials. FUNDING: US National Institutes of Health, St Baldrick's Foundation, and Alex's Lemonade Stand Foundation.

Comparative effectiveness of high-power holmium laser lithotripsy for pediatric patients with kidney and ureteral stones.

BACKGROUND: The comparative effectiveness of high-power laser technology for kidney stone surgery in pediatric patients is poorly understood. We compared outcomes for the 120 W Holmium:yttrium-aluminum-garnet (Ho:YAG) laser with MOSES technology to 30 W Ho:YAG laser for pediatric patients undergoing ureteroscopy with laser lithotripsy for kidney and ureteral stones. OBJECTIVE: We evaluated the outcomes of the new MOSES laser technology as compared to low-power Ho:YAG lasers commonly used for kidney stone treatment in the pediatric population. METHODS: We performed a retrospective cohort study of 131 consecutive patients aged 1-18 years who underwent ureteroscopy and laser lithotripsy for renal and ureteric calculi at a large freestanding children's hospital between 2013 and 2020. The primary outcome was the efficiency quotient, which incorporates stone clearance, auxiliary procedures, and retreatment rates. Outcomes were compared between groups using Chi-square or Fisher's exact tests and multivariable regression. A sensitivity analysis was performed extending the age limit to ≤21 years. RESULTS: Outcomes are summarized in the table below. Median age of the cohort was 14 years with 53% of patients being female. MOSES laser had a higher efficiency quotient and was associated with a lower odds of post-operative emergency department visits (OR 0.2, 95% CI 0.0-1.0; p = 0.047). Operative time was similar. In the sensitivity analysis of patients ≤21 years, the statistical significance with fewer emergency department visits was lost and the efficiency quotient was lower. DISCUSSION: Our results show that stone clearance is similar between the 120 W MOSES and 30 W Ho:YAG lasers. However, there are indications that high-power laser lithotripsy is more efficient due to fewer auxiliary procedures and a reduction in retreatment. In addition, higher power lasers were associated with fewer emergency department visits. The benefits appear to be greater among children ≤18 years. These exploratory findings are important for pediatric patients due to the requirement for general anesthesia for each procedure and their associated impact on children and their caregivers. CONCLUSIONS: High-power laser lithotripsy may be more efficient than lower power laser lithotripsy, which is driven by the fewer auxiliary procedures and reduction in retreatment particularly among youth ≤18 years old.

Longitudinal Large-Scale Semiquantitative Proteomic Data Stability Across Multiple Instrument Platforms.

With the rapid developments in mass spectrometry (MS)-based proteomics methods, label-free semiquantitative proteomics has become an increasingly popular tool for profiling global protein abundances in an unbiased manner. However, the reproducibility of these data across time and LC-MS platforms is not well characterized. Here, we evaluate the performance of three LC-MS platforms (Orbitrap Elite, Q Exactive HF, and Orbitrap Fusion) in label-free semiquantitative analysis of cell surface proteins over a six-year period. Sucrose gradient ultracentrifugation was used for surfaceome enrichment, following gel separation for in-depth protein identification. With our established workflow, we consistently detected and reproducibly quantified >2300 putative cell surface proteins in a human acute myeloid leukemia (AML) cell line on all three platforms. To our knowledge this is the first study reporting highly reproducible semiquantitative proteomic data collection of biological replicates across multiple years and LC-MS platforms. These data provide experimental justification for semiquantitative proteomic study designs that are executed over multiyear time intervals and on different platforms. Multiyear and multiplatform experimental designs will likely enable larger scale proteomic studies and facilitate longitudinal proteomic studies by investigators lacking access to high throughput MS facilities. Data are available via ProteomeXchange with identifier PXD022721.

Structural mechanism of DNA recognition by the p202 HINa domain: insights into the inhibition of Aim2-mediated inflammatory signalling.

The HIN-200 family of proteins play significant roles in inflammation-related processes. Among them, AIM2 (absent in melanoma 2) and IFI16 (γ-interferon-inducible protein 16) recognize double-stranded DNA to initiate inflammatory responses. In contrast, p202, a mouse interferon-inducible protein containing two HIN domains (HINa and HINb), has been reported to inhibit Aim2-mediated inflammatory signalling in mouse. To understand the inhibitory mechanism, the crystal structure of the p202 HINa domain in complex with a 20 bp DNA was determined, in which p202 HINa nonspecifically recognizes both strands of DNA through electrostatic attraction. The p202 HINa domain binds DNA more tightly than does AIM2 HIN, and the DNA-binding mode of p202 HINa is different from that of the AIM2 HIN and IFI16 HINb domains. These results, together with the reported data on p202 HINb, lead to an interaction model for full-length p202 and dsDNA which provides a conceivable mechanism for the negative regulation of Aim2 inflammasome activation by p202.

Structural basis for the regulation of maternal embryonic leucine zipper kinase.

MELK (maternal embryonic leucine zipper kinase), which is a member of the AMPK (AMP-activated protein kinase)-related kinase family, plays important roles in diverse cellular processes and has become a promising drug target for certain cancers. However, the regulatory mechanism of MELK remains elusive. Here, we report the crystal structure of a fragment of human MELK that contains the kinase domain and ubiquitin-associated (UBA) domain. The UBA domain tightly binds to the back of the kinase domain, which may contribute to the proper conformation and activity of the kinase domain. Interestingly, the activation segment in the kinase domain displays a unique conformation that contains an intramolecular disulfide bond. The structural and biochemical analyses unravel the molecular mechanisms for the autophosphorylation/activation of MELK and the dependence of its catalytic activity on reducing agents. Thus, our results may provide the basis for designing specific MELK inhibitors for cancer treatment.