AI News, Key challenges for delivering clinical impact with artificial intelligence artificial intelligence

A principle tenet of precision medicine is that subpopulations may be reasonably identified who differ in their disease risk, prognosis and response to treatment due to differences in underlying biology and other characteristics.

The availability of multidimensional datasets that capture such variation can be ‘trained’ using artificial learning algorithms to find the cryptic phenotypic or genotypic structures, discussed subsequently, to then predict risk of disease, treatment response, prognosis and other outcomes in individual patients based on their own characteristics.

The realization of this will offer clinicians the opportunity to more carefully tailor interventions—whether disease modifying or preventative in nature—to individual patients, contrasting with the current inductive process of symptom classification, and sometimes vague and inexact process of treatment decisions.

However, computer capabilities have grown exponentially in recent years, and the integrated efforts of the international scientific community have made available large multidimensional biological and clinical datasets.1,2,3,4,5 Recently, prediction algorithms utilizing artificial intelligence approaches for cancer6,7,8,9 and cardiovascular disease10,11 have shown promising results, predicting disease risk with a higher degree of precision.

Realizing a similar approach to the group of disorders of brain development termed ‘neurodevelopmental disorders’ (NDD) has a number of obstacles.12 The NDDs are a group of early childhood onset disorders that impact different domains of cognitive development, motor function and other higher brain functions, and are lifelong in nature.

Among the NDDs are severe disorders that impact multiple domains of cognitive functioning, such as intellectual disability (ID), as well as severe and pervasive disorders of social communication (autism spectrum disorder, ASD), motor function and cognition (epilepsy encephalopathies), and behavioral regulation (attention deficit hyperactivity disorder, ADHD).

As such, early diagnosis and targeted therapeutic interventions to those who are most likely to benefit are universally agreed public health priorities.16,17 In this review, the ambition of precision medicine will be described, and success and implementation in medical practice so far will be briefly presented, with certain cancers and cardiovascular disease as examples of success.

Deep learning algorithm to diagnose heart attack using 549 ECG records shows a sensitivity of 93.3% and specificity of 89.7%, outperforming cardiologists.28 Recently, DNA sequencing technology adopted machine learning to read out long stretches of DNA fragments from digital electronic signaling data.

This method has accuracy over 98% and can produce mega base long DNA reads.29 There has been an attempt to use AI in the clinical classification of genomic variation, based on the characterization of non-coding variants30 splicing code,5 DNA/RNA binding proteins2 and non-coding RNA (ncRNA)31 using large-scale molecular datasets.

Classifying mutations according to their clinical relevance is very complex due to the largely unknown penetrance of individual variants, (i.e., the probability of diagnosis given a particular variant is identified, or mathematically, P(disease+|variant+)) Moreover, high penetrance variants are largely infrequent, with those of low penetrance much more common.

Despite this, however, recent deep learning methods have had some degree of success in the correct interpretation of phenotype and genomic data for disease risk in numerous types of cancer,6,7,9,32,33 diabetic retinopathy34,35 and pharmacogenomics.36,37,38 For example, in discriminating lymph node metastases, 7 independent deep learning implementations showed greater discrimination power (i.e., in relation to pathological versus non-pathological) compared to 11 pathologists.7 The best deep learning algorithm performed with an area under the curve (AUC) of 0.99, compared to 0.88 for ‘best’ clinician-derived score.

Neurodevelopmental disorders have their onset early in childhood and impact on a variety of functional domains, including cognition and executive function, language and social function, and motor function and behavior control.39,40,41 A number of different diagnoses are subsumed within this category, including intellectual disability (ID),42 autism spectrum disorder (ASD),4,43 attention deficit hyperactivity disorder (ADHD),44 tic disorders, and other movement disorders.45,46 Epilepsy and other early onset brain disorders, with or without associated ID, are also classified as NDDs47,48 (Table 1).

All the NDDs considered in this current discussion are principally genetic in etiology.50 For example, the early twin and family studies in ASD all supported a strong, heritable genetic component, and ASD and a lesser phenotype termed the Broader Autism Phenotype (BAP) do tend to run in families.51 Some individual cases may result from rare, highly penetrant mutations,4,43,52 some of which segregate in a Mendelian fashion.

In contrast, however, most appear to result from a more complex genetic architecture that involves one or more genetic variants of variable penetrance interacting with other epigenetic mechanisms and environmental factors.13,55,56 Understanding this genetic complexity is important, and it is anticipated that technological developments, both in silico but also laboratory based, will help unravel this.

To date, over 250 genes have been reported to have strong association with NDD.58 A very small number of genes (SCN2A, CHD8, STXBP1) and loci (16p11.2 microdeletion, 15q13.3 microdeletion etc.) that are found to be enriched within NDD are still below the level of 1% frequency threshold.48,59,60,61 The current clinical genetic diagnostic yield for severe, syndromic NDDs associated with ID is approximately 40% and it is higher if genome sequencing data are available for other members (parents, siblings) of the family.62 In imaging studies, similarities in brain function evident from fMRI and diffusion tensor imaging also point to overlap at the level of intermediate phenotype between a number of NDDs such as ASD and ADHD.63,64 Studies have examined diagnosed individuals while performing different neuropsychological tasks in the scanner, and the regions and structures in the brain that are active have been elucidated.

As we discuss subsequently, machine learning offers the opportunity to examine biological datasets in both a supervised and unsupervised manner, thereby providing both predictive models for diagnosis and treatment, as well as, theoretically, examining how multidimensional datasets may inform new models of classification.

Since then, the human genome has been mapped67,68 and exome and whole-genome sequencing technologies have led to the detection of hundreds of disease causal genes and loci for ASD and other NDDs.43,69,70,71 Indeed, conducting exome or genome sequencing for newborn babies at high risk of genetic abnormalities is now becoming more frequent and cost effective.72 Subsequently, the advent of transcriptome sequencing dependent technologies led to the establishment of the Allen developmental human brain atlas73 in 2011, ENCODE database profiling the non-coding elements in the human genome74 in 2012, and the Human Cell Atlas75 in 2017.

Multiple sequencing consortiums focussed on the NDDs were also started during the period of 2012 and 2014 with the aim of identifying disease-implicated variants, and making exome and WGS data available to the scientific community for further study.52,70,76,77 Bearing in mind that most identified genetic variation is of unknown pathogenicity, and little is known about functional consequences, the discovery of CRISPR/Cas as a gene editing tool in 2012 has allowed scientists to better characterize identified genetic variants.78,79 In recent years, artificial intelligence approaches has been used in autism spectrum disorder,5,12,14,15 epileptic encephalopathy,80,81,82 intellectual disability,83,84,85 attention deficit hyperactivity disorder (ADHD),86 and rare genetic disorders.2 In our discussion of AI in NDDs, three layers of analyses will be considered.

Although the current genetic diagnostic yield (including copy number variation (CNV), single nucleotide variants (SNV), and indel) for severe, syndromal ID is around 50% (Table 1) we still do not know genes or loci for NDDs more generally, which includes many of the cases whether there is no ID and/or evidence of craino-facial dysmorphology.58,62 In addition, many identified loci are confounded by unknown penetrance, and, beyond bioinformatic prediction, do not have a strong evidential basis of support.

For example, 15q13.3 microdeletion syndrome impacts multiple domains of cognitive function and is associated with heterogenous phenotypes, including epilepsy/seizure (57%), speech delay (16%), and ASD (11%).90 There are hundreds of such CNVs with no straightforward mapping between manifested phenotypes and the variants/genes.40,58 Despite the possibilities, there remains the problem of phenotype, and in particular, the oversimplification of dichotomizing phenotypes such as ASD and ADHD into ‘caseness’.

ASD, that may not correctly capture the structure in the underlying data.92 Similarly, in epilepsy, EEG endophenotypes have been proposed93 and purely EEG-based classification of seizures have been investigated theoretically and clinically.94 However, none of these methods have been applied in a quantitative context, perhaps as the diagnosis of subtypes of epilepsy often rests heavily on qualitative EEG observations.

A neural network based approach on quantifying gene score for polygenic trait (i.e height) using single nucleotide polymorphism data showed promising improvements out performing previous methods.95 Polygenic risk prediction in NDDs remains problematic in light of the largely negative findings from underpowered genome-wide association studies (GWAS).

Recently, a genome wide association study on large autism spectrum disorder and control cohort identified five common variants that confers very small risk factor.57 When the proportion from de novo risk factor is substantially large, it is still not clear to what extent common variants contributes into the genetic risk factor of neurodevelopmental disorders.

Statistical significance for the large number of interactions also suffers from the impact of multiple testing thresholds.96,97 Although gene–gene interaction is likely a major contributor to the phenotypic variance of NDDs, there is currently no credible artificial intelligence algorithm able to cope with data on this scale.

Certainly, a large number of genes can be simplified into a smaller number of protein–protein interactions or co-expression networks using traditional statistical model or algorithms, but as discussed subsequently, this is computationally NP-hard.98 Adding to this complexity, there may be significant overlaps between gene lists and/or protein/co-expression networks for different neurodevelopmental disorders, and so discriminatory classification of, say, ASD vis-à-vis schizophrenia adds further layers of complexity.

For example, mTOR pathway impacts a certain group of epilepsy individuals and the same pathway found to be dysregulated in tuberous sclerosis individuals.54,100 Hence, mTOR inhibitors have a great potential to impact treatment outcome for individuals with epilepsy carrying mTOR mutation or tuberous sclerosis related epilepsy.

For example, reliability and reproducibility of neuroimaging findings depend hugely on many experimental factors.103 Similarly, population scale omics data suffer from batch effect and technology specific biases.104,105 Thus, although large databases may be available for machine learning approaches, great care has to be taken in the quality and comparability of datasets used.

For example, there is evidence of certain somatic mutations associated with autism spectrum disorder, microcephaly, and epilepsy.48,60,88 Recent analysis has also shown that up to 40% of neurons could have a large mega base scale copy number variation.109 Single cell genomics has also identified private somatic mutations within each neuron.89 Although the contribution of somatic mutation to disease risk is not well understood, this particular type of mutation will add unpredictable variance within machine learning approaches and will impact replication significantly.

For example, in genetic algorithms, it can be extremely difficult for clinicians to decipher how through random operations (i.e., mutation, crossover) and variables the model reaches fitness convergence for optimum solutions in a multidimensional search space.113 Ultimately, however, despite the complexity of different algorithms, statistical models and tests are used to favor or refute evidence (i.e., p-values, false discovery rates, area under the curve), which can be understood by many professionals working in a clinical setting.114 The datasets themselves, multidimensional in nature, will also have been collected from multidisciplinary experts who may not necessarily ‘talk the same language’.

Regarding the delivery of genome editing machineries, recent efforts on vector and non-vector based CRISPR system delivery shows limited success on breaking the blood-brain barrier.116,117 For NDD, future treatment options should implement AI based algorithms that can design genome editing or antisense oligonucleotide design tools that are compatible with the in vivo delivery mechanism.

Natural language processing (NLP) is another emerging field of machine intelligence that can automatically transform clinical text into structured clinical data.118 NLP algorithms can analyze digital health records and psychiatric notes to identify relatedness among patients’ phenotypes and their associated genetic markers.

The challenge is to understand the manifestation of wide ranges of phenotypes during developmental stages in an individual that arises from the same genetic and neuronal substrate.39,47,58 Application of artificial intelligence algorithm on longitudinal studies can be designed to capture the pattern of disease progression over time and the variability at the personal or sub-population level.

Artificial intelligence in clinical and genomic diagnostics

Mimicking human intelligence is the inspiration for AI algorithms, but AI applications in clinical genomics tend to target tasks that are impractical to perform using human intelligence and error prone when addressed with standard statistical approaches.

Many of the techniques described above have been adapted to address the various steps involved in clinical genomic analysis—including variant calling, genome annotation, variant classification, and phenotype-to-genotype correspondence—and perhaps eventually they can also be applied for genotype-to-phenotype predictions.

In addition, recent results suggest that deep learning is poised to revolutionize base calling (and as a result, variant identification) for nanopore-based sequencing technologies, which have historically struggled to compete with established sequencing technology because of the error-prone nature of prior base-calling algorithms [55].

Here, we describe both genome annotation and variant classification because many of the AI algorithms that are used to predict the presence of a functional element from primary DNA sequence data are also used to predict the impact of a genetic variation on those functional elements.

Some of these methods have been integrated into deep-learning-based meta-predictors (models that process and merge the predictions produced by several other predictors) that outperform both their individual predictive components and the combination of those predictive components when integrated using regression or other machine-learning approaches [57].

In addition, deep generative models have shown promise for predicting the effects of genetic variants [61], and are especially intriguing given their ability to evaluate the joint influence of multiple genetic variants and/or complex indels on protein function, a capability that is largely absent from most pathogenicity prediction tools.

Splicing defects in genes are responsible for at least 10% of rare pathogenic genetic variation [63], but they can be difficult to identify because of the complexity of intronic and exonic splicing enhancers, silencers, insulators, and other long range and combinatorial DNA interactions that influence gene splicing [64].

Remarkably, SpliceAI was able to use long-range sequence information to boost prediction accuracy from 57%, using a short window size (80 nucleotides) typical for many prior splicing prediction tools, to 95% when a 10 kb window size was ingested by the AI algorithm, and was able to identify candidate cryptic splicing variants underlying neurodevelopmental disorders.

DeepSEA, a multitask hierarchically structured CNN trained on large-scale functional genomics data [67], was able to learn sequence dependencies at multiple scales and simultaneously produce predictions of DNase hypersensitive sites, transcription factor binding sites, histone marks, and the influence of genetic variation on those regulatory elements, with a level of accuracy superior to those of other tools for prioritizing non-coding functional variants [68].

DeepGestalt, a CNN-based facial image analysis algorithm, dramatically outperforms human dysmorphologists in this task and is precise enough to distinguish between molecular diagnoses that are mapped to the same clinical diagnosis (that is, distinct molecular forms of Noonan syndrome) [19].

When combined with genomic data, PEDIA, a genome interpretation system incorporating DeepGestalt, was able to use phenotypic features extracted from facial photographs to accurately prioritize candidate pathogenic variants for 105 different monogenic disorders across 679 individuals [74].

A ‘survival CNN’, which is a combination of a CNN with Cox proportional hazards-based outcomes (a type of statistical survival analysis), was able to learn the histological features of brain tumors that are associated with survival and correlated with somatic mutation status [75].

More generally, AI-based computer vision systems appear to be capable of predicting the genomic aberrations that are likely to be present in an individual’s genome on the basis of the complex phenotypes embedded in relevant clinical images [20, 75].

A hierarchical statistical model, tiered on the basis of anatomic divisions in a manner meant to mimic the clinical reasoning of a composite of experienced physicians, was trained on the NLP output to generate a diagnostic system [48].

For example, an NLP system was designed to extract phenotypic descriptions automatically from EHR data of pediatric patients with rare diseases, and to rank matches to the expected phenotypic features of candidate pathogenic variants in the patients’ genomes [49].

A few studies have attempted genomic prediction of complex human traits using AI algorithms, but most of those reported in the literature to date are probably overfit as they purportedly explain substantially more trait variance than should be possible on the basis of heritability estimates.

How is AI being applied to healthcare challenges?

This video was shown during Global Breakthrough Day by Philips HealthWorks on December 12, 2018 at the Evoluon in Eindhoven.

2018 WMIF | AI and the Cost of Trials: The Impact of Real World and Real Time Evidence

Moderator: Krishna Yeshwant, MD, Partner, Google Ventures; Instructor in Medicine, BWH Amy Abernethy, MD, PhD, CMO, CSO & SVP Oncology, Flatiron ...

Towards ambient intelligence in AI-assisted healthcare spaces - Dr Fei-Fei Li, Stanford University

Abstract: Artificial intelligence has begun to impact healthcare in areas including electronic health records, medical images, and genomics. But one aspect of ...

AI and Machine Learning in Health Programs

After viewing the webinar recording, please fill out our webinar evaluation form: ...

Empowering patients with technology in clinical trials

Here from Joe Dustin, Principal of Mobile Health Health chat about delivering a patient-centric approach for clinical trials: Current trends and new opportunities.

Data Science and AI in Pharma and Healthcare (CXOTalk #275)

Data, artificial intelligence and machine learning are having a profound influence on healthcare, drug discovery, and personalized medicine. On this episode ...

AI for Healthcare Leaders: The New AI Frontier for Improved Leadership Decision Making

A new frontier is expanding AI from artificial intelligence to augmented intelligence. Traditional AI has focused on improving analytics efficiency and effectiveness ...

Bringing AI and machine learning innovations to healthcare (Google I/O '18)

Could machine learning give new insights into diseases, widen access to healthcare, and even lead to new scientific discoveries? Already we can see how ...

#213: Artificial Intelligence and the Digital Healthcare Revolution

213: AI and the Digital Healthcare Revolution Artificial intelligence is disrupting healthcare, from the largest institutions to the most intimate doctor-patient ...

Google AI Impact Challenge: Using Technology to Change the World (Google I/O'19) President Jacquelline Fuller talks to global social innovators being supported through the Google AI Impact Challenge who are using AI to address ...