AI-Informed Management of Children in ICUs Stephen Kingsmore, MD - PowerPoint PPT Presentation

AI-Informed Management of Children in ICUs Stephen Kingsmore, MD DSc

No conflict of interest with regard to this presentation et al

17 month old boy with fever (T max 103 0 F) x 3 days, vomiting, diarrhea, abdominal pain, labored breathing, skin lesions x 1 day

Rady Emergency Department Work Up • Blood tests: • Metabolic panel: metabolic acidosis • C-reactive protein: markedly elevated • Complete blood count: low white cell count • Abdominal ultrasound & computed tomography: No intussusception, possible mild colitis • Lumbar puncture • Cardiovascular decompensation = hypovolemic shock → intravenous fluids • SiO 2 88% on Fi0 2 21% → Continuous Positive Airway Pressure ventilation • Sepsis suspected → intravenous vancomycin + ceftriaxone • Admitted to PICU → switched to intravenous meropenem

Hospital Day 2 • Blood culture: Pseudomonas aeruginosa • Skin rash diagnosed as echthyma gangrenosum • AI-Informed Management ordered

AI-Informed Management of Critically Ill Children • 30,000 rare or ultra-rare genetic diseases • Cause ~15% of admissions to level IV neonatal intensive care units & leading cause of infant mortality • Specific treatments are available for many Weeks of Traditional empiric Management treatment Search for Improvement or 8% Genetic 1% Precision 0% Change in etiological worsening Disease Diagnosis Medicine Outcome diagnosis Treatment Modification Critically Etiologic ill child diagnosis admitted 24 hours of unknown empiric to ICU treatment Ultra-Rapid 47% Genetic ~20% Change Genome 89% Actionable Disease Diagnosis in Outcome Sequencing AI-Informed Search for ~33% Rule-out etiological Specific Genetic Management diagnosis Diseases Refined Differential Diagnosis

AI-driven diagnosis of genetic diseases in children in ICUs urWGS ordered on day of admission with 1-2 day time to result is optimal in order to change the care and outcomes of these critically ill neonates and children 7 Farnaes et al 2018, Willig et al 2015, Petrikin et al 2018, Clark et al 2019

Step 1: Order in Epic Electronic Health Record Pseudomonal sepsis, leukopenia Yes

Step 2: Sample Preparation (2.5 hours) • 0.5 ml blood • Illumina Nextera Flex Fragmentation No PCR amplification

Step 3: Genome Sequencing (15.5 hours) • 2 x 100 nucleotide paired sequences • Illumina NovaSeq 6000 instrument • S1 flowcell • Trio or 2 Probands per flowcell • 40X proband; 30X parents

Step 4: Identify all disease-causing variants in child’s genome: 45 min 125 billion nucleotides sequenced 125 billion nucleotides sequenced 125 billion nucleotides sequenced 2.8 billion genomic nucleotides 2.8 billion genomic nucleotides 2.8 billion genomic nucleotides assigned assigned assigned 4.9 million variants 4.9 million variants 4.9 million variants identified & genotyped identified & genotyped identified & genotyped Glossary: Nucleotide – a single DNA letter (base); Adenine, Cytosine, Guanine or Thymidine Variant – a DNA change from the (normal) reference genome sequence Illumina DRAGEN 2.0

Step 5: Variant pathogenicity scoring Very Null variant (nonsense, frameshift, ±1 or 2 splice site position, initiation codon, Variant Category Criteria Strong (VS) exon deletion) in gene where LOF known to cause disease • Strong (S) Same amino acid change as previously established pathogenic variant Pathogenic (P): 99% 1 VS + (1S or • De novo in a patient with the disease and no family history disease causing 2M/Sup) • Functional studies show damaging effect on the gene 2S • Prevalence in affected individuals significantly greater than controls 1S + (3M or • Moderate Located in mutational hot spot/functional domain without benign variation 2M+2Supp) • (M) Extremely low frequency in Gnomad • Recessive disorders, detected in trans with a pathogenic variant Likely Pathogenic 1 VS/S + 1 M • Protein length changed by in-frame indel in nonrepeat region or stop-loss (LP): 90% disease 1 S + (1 M or 2 • Novel missense at amino acid where different missense known to be causing Supp) pathogenic 3 M • Assumed de novo, but without confirmation of paternity and maternity 2 M + 2 Supp • Supporting Cosegregation with disease in multiple affected family members in gene 1 M + 4 Supp (Supp) known to cause disease • Missense variant in gene with low rate of benign missense variants and Variant of Uncertain where missense variants commonly cause disease Significance (VUS): • Multiple computational tools call deleterious 10% disease • Phenotype highly specific for disease with single genetic etiology causing • Reputable source reports as pathogenic, but unpublished Richards S, et al. Standards and guidelines for the interpretation of sequence variants: a joint consensus recommendation of ACMG and AMP. Genet Med. 2015 Mar 5.

Step 5: Variant Pathogenicity Scoring: 2 mins 125,100,000,000 2,800,000,000 4,872,577 711,870 e l p o e p 0 0 1 : 1 < n i t n e s e r p s t n a i r a 962 V s t n a r i a V c i n e g o h t a P y l e k L i d n a c n i e g o h t a P Diploid MOON software with InterVar post-processing

Why collect a deep phenotype • The clinical features of NICU Mean patient infants do NOT correspond well phenotypes in EHR: 93.1 Mean patient with classical descriptions of their phenotypes disease provided by MD: 5.0 • The ability to make a diagnosis is critically dependent on a full clinical description Mean disease phenotypes in text book: 93.1 Glossary: Phenotype – the clinical features of a patient with a disease 76 children with genetic diseases; natural language processing of EHR; Text book: Mendelian Inheritance in Man

Step 6: Deep Phenotyping by Natural Language Processing of Epic EMR: 20 sec CliniThink CLiXENRICH natural language processing software

Step 7: Translate phenotypes to a hierarchical standardized vocabulary

Step 8: Pattern Recognition creates a comprehensive differential diagnosis

Step 9: Automated Diagnosis: 2 mins Manual Diagnosis: 1 – 10 hours 125,100,000,000 2,800,000,000 4,872,577 711,870 s e s a e s i D c t i e n e G x 14,000 s e p y t o n e h P 962 x 159 s t n a r i a V c i n e g o h t a P y l e k L i d n a c 1 n i e g o h s t s i a o P n g a d i l a n o s i i v o r P Diploid MOON software with InterVar post-processing

Step 9: Automated Diagnosis: 4 mins Manual Diagnosis: 10 hours • X-linked recessive • Inherited from mother • Loss of splice donor site of intron 11 • Classified as pathogenic • Confirmed by functional studies

Hospital Course • Diagnosis after 22 hours • Individualized medicine • Double coverage, double duration antibiotics • Intravenous immunoglobulin to maintain IgG level >600mg/dL • Magnetic resonance imaging: no additional septic emboli • Prognosis • Normal life • 10% have significant infections despite treatment • Genetic counseling • Mother is a carrier • Maternal relatives at-risk • Discharged home on day 13 Smith and Berglof, Gene Reviews, 2016

Diagnostic performance of 3 rd generation rWGS-based individualized medicine • Retrospective, n=84 children with 86 diagnoses • Expert manual interpretation: Precision 98%, Recall 98% • Automated interpretation: Precision 99%, Recall 95% Retrospective Patients Prospective Patients Subject ID 263 6124 3003 6194 290 352 362 374 7052 412 Age 8 days 14 years 1 year 5 days 3 days 7 weeks 4 weeks 2 days 17 months 3 days Abbreviated Neonatal Rhabdo- Dystonia, Hypoglycemia Pulmonary Diabetic Neonatal Pseudomonal Neonatal HIE, anemia Presentation seizures myolysis Dev. delay seizures hemorrhage ketoacidosis seizures septic shock seizures Method Auto. Auto. Auto. Auto. Auto. Std. Auto. Std. Auto. Std. Auto. Std. Auto. Std. Auto. Std. Auto. Std. Total (hours) 20:25 19:56 19:20 19:14 20:42 * 56:03 19:29 48:46 19:11 42:04 19:10 57:21 31:02 † 34:38 22:04 38:37 20:53 48:23 Early Infantile Glycogen Dopa- Permanent X-linked Benign familial Molecular None None None None Epileptic Storage Responsive neonatal agamma- neonatal Diagnosis Encephalo- Disease V Dystonia diabetes globulinemia seizures 1 Gene and PYGM TH KCNQ2 BTK KCNQ2 Causative c.2262delA c.785C>G n.a. n.a. INS c.26C>G n.a. n.a. c.727C>G c.974+2T>C c.1051C>G Variant(s) c.1726C>T c.541C>T

Scaling to meet national need 1,200 NICUs in 30 countries working to continuously improve neonatal care Our Signature Site 22 Legend: = Active sites = Sites pending formal agreements

Summary: Genomic Medicine will be the first AI-driven specialty • Technology is evolving faster than we can train physicians • Broad implementation of genomic medicine in children will be dependent on AI for 4 reasons: • There are 30,000 genetic diseases – too many for physician computation • There are a dearth of clinical trained genomic medicine practitioners • Disease progression in newborns is often too rapid for traditional approaches • Implementation and clinical trials of new therapies are co-occurring