Artificial intelligence in epilepsy - applications and pathways to the clinic.

Artificial intelligence (AI) is rapidly transforming health care, and its applications in epilepsy have increased exponentially over the past decade. Integration of AI into epilepsy management promises to revolutionize the diagnosis and treatment of this complex disorder. However, translation of AI into neurology clinical practice has not yet been successful, emphasizing the need to consider progress to date and assess challenges and limitations of AI. In this Review, we provide an overview of AI applications that have been developed in epilepsy using a variety of data modalities: neuroimaging, electroencephalography, electronic health records, medical devices and multimodal data integration. For each, we consider potential applications, including seizure detection and prediction, seizure lateralization, localization of the seizure-onset zone and assessment for surgical or neurostimulation interventions, and review the performance of AI tools developed to date. We also discuss methodological considerations and challenges that must be addressed to successfully integrate AI into clinical practice. Our goal is to provide an overview of the current state of the field and provide guidance for leveraging AI in future to improve management of epilepsy.

Network coupling and surgical treatment response in temporal lobe epilepsy: A proof-of-concept study.

OBJECTIVE: This proof-of-concept study aimed to examine the overlap between structural and functional activity (coupling) related to surgical response. METHODS: We studied intracranial rest and ictal stereoelectroencephalography (sEEG) recordings from 77 seizures in thirteen participants with temporal lobe epilepsy (TLE) who subsequently underwent resective/laser ablation surgery. We used the stereotactic coordinates of electrodes to construct functional (sEEG electrodes) and structural connectomes (diffusion tensor imaging). A Jaccard index was used to assess the similarity (coupling) between structural and functional connectivity at rest and at various intraictal timepoints. RESULTS: We observed that patients who did not become seizure free after surgery had higher connectome coupling recruitment than responders at rest and during early and mid seizure (and visa versa). SIGNIFICANCE: Structural networks provide a backbone for functional activity in TLE. The association between lack of seizure control after surgery and the strength of synchrony between these networks suggests that surgical intervention aimed to disrupt these networks may be ineffective in those that display strong synchrony. Our results, combined with findings of other groups, suggest a potential mechanism that explains why certain patients benefit from epilepsy surgery and why others do not. This insight has the potential to guide surgical planning (e.g., removal of high coupling nodes) following future research.

The seizure severity score: a quantitative tool for comparing seizures and their response to therapy.

Objective.Epilepsy is a neurological disorder characterized by recurrent seizures which vary widely in severity, from clinically silent to prolonged convulsions. Measuring severity is crucial for guiding therapy, particularly when complete control is not possible. Seizure diaries, the current standard for guiding therapy, are insensitive to the duration of events or the propagation of seizure activity across the brain. We present a quantitative seizure severity score that incorporates electroencephalography (EEG) and clinical data and demonstrate how it can guide epilepsy therapies.Approach.We collected intracranial EEG and clinical semiology data from 54 epilepsy patients who had 256 seizures during invasive, in-hospital presurgical evaluation. We applied an absolute slope algorithm to EEG recordings to identify seizing channels. From this data, we developed a seizure severity score that combines seizure duration, spread, and semiology using non-negative matrix factorization. For validation, we assessed its correlation with independent measures of epilepsy burden: seizure types, epilepsy duration, a pharmacokinetic model of medication load, and response to epilepsy surgery. We investigated the association between the seizure severity score and preictal network features.Main results.The seizure severity score augmented clinical classification by objectively delineating seizure duration and spread from recordings in available electrodes. Lower preictal medication loads were associated with higher seizure severity scores (p= 0.018, 97.5% confidence interval = [-1.242, -0.116]) and lower pre-surgical severity was associated with better surgical outcome (p= 0.042). In 85% of patients with multiple seizure types, greater preictal change from baseline was associated with higher severity.Significance.We present a quantitative measure of seizure severity that includes EEG and clinical features, validated on gold standard in-patient recordings. We provide a framework for extending our tool's utility to ambulatory EEG devices, for linking it to seizure semiology measured by wearable sensors, and as a tool to advance data-driven epilepsy care.

Remote effects of temporal lobe epilepsy surgery: Long-term morphological changes after surgical resection.

OBJECTIVE: Epilepsy surgery is an effective treatment for drug-resistant patients. However, how different surgical approaches affect long-term brain structure remains poorly characterized. Here, we present a semiautomated method for quantifying structural changes after epilepsy surgery and compare the remote structural effects of two approaches, anterior temporal lobectomy (ATL), and selective amygdalohippocampectomy (SAH). METHODS: We studied 36 temporal lobe epilepsy patients who underwent resective surgery (ATL = 22, SAH = 14). All patients received same-scanner MR imaging preoperatively and postoperatively (mean 2 years). To analyze postoperative structural changes, we segmented the resection zone and modified the Advanced Normalization Tools (ANTs) longitudinal cortical pipeline to account for resections. We compared global and regional annualized cortical thinning between surgical treatments. RESULTS: Across procedures, there was significant cortical thinning in the ipsilateral insula, fusiform, pericalcarine, and several temporal lobe regions outside the resection zone as well as the contralateral hippocampus. Additionally, increased postoperative cortical thickness was seen in the supramarginal gyrus. Patients treated with ATL exhibited greater annualized cortical thinning compared with SAH cases (ATL: -0.08 ± 0.11 mm per year, SAH: -0.01 ± 0.02 mm per year, t = 2.99, P = 0.006). There were focal postoperative differences between the two treatment groups in the ipsilateral insula (P = 0.039, corrected). Annualized cortical thinning rates correlated with preoperative cortical thickness (r = 0.60, P < 0.001) and had weaker associations with age at surgery (r = -0.33, P = 0.051) and disease duration (r = -0.42, P = 0.058). SIGNIFICANCE: Our evidence suggests that selective procedures are associated with less cortical thinning and that earlier surgical intervention may reduce long-term impacts on brain structure.

Quantitative approaches to guide epilepsy surgery from intracranial EEG.

Over the past 10 years, the drive to improve outcomes from epilepsy surgery has stimulated widespread interest in methods to quantitatively guide epilepsy surgery from intracranial EEG (iEEG). Many patients fail to achieve seizure freedom, in part due to the challenges in subjective iEEG interpretation. To address this clinical need, quantitative iEEG analytics have been developed using a variety of approaches, spanning studies of seizures, interictal periods, and their transitions, and encompass a range of techniques including electrographic signal analysis, dynamical systems modeling, machine learning and graph theory. Unfortunately, many methods fail to generalize to new data and are sensitive to differences in pathology and electrode placement. Here, we critically review selected literature on computational methods of identifying the epileptogenic zone from iEEG. We highlight shared methodological challenges common to many studies in this field and propose ways that they can be addressed. One fundamental common pitfall is a lack of open-source, high-quality data, which we specifically address by sharing a centralized high-quality, well-annotated, multicentre dataset consisting of >100 patients to support larger and more rigorous studies. Ultimately, we provide a road map to help these tools reach clinical trials and hope to improve the lives of future patients.

Spike patterns surrounding sleep and seizures localize the seizure-onset zone in focal epilepsy.

OBJECTIVE: Interictal spikes help localize seizure generators as part of surgical planning for drug-resistant epilepsy. However, there are often multiple spike populations whose frequencies change over time, influenced by brain state. Understanding state changes in spike rates will improve our ability to use spikes for surgical planning. Our goal was to determine the effect of sleep and seizures on interictal spikes, and to use sleep and seizure-related changes in spikes to localize the seizure-onset zone (SOZ). METHODS: We performed a retrospective analysis of intracranial electroencephalography (EEG) data from patients with focal epilepsy. We automatically detected interictal spikes and we classified different time periods as awake or asleep based on the ratio of alpha to delta power, with a secondary analysis using the recently published SleepSEEG algorithm. We analyzed spike rates surrounding sleep and seizures. We developed a model to localize the SOZ using state-dependent spike rates. RESULTS: We analyzed data from 101 patients (54 women, age range 16-69). The normalized alpha-delta power ratio accurately classified wake from sleep periods (area under the curve = .90). Spikes were more frequent in sleep than wakefulness and in the post-ictal compared to the pre-ictal state. Patients with temporal lobe epilepsy had a greater wake-to-sleep and pre- to post-ictal spike rate increase compared to patients with extra-temporal epilepsy. A machine-learning classifier incorporating state-dependent spike rates accurately identified the SOZ (area under the curve = .83). Spike rates tended to be higher and better localize the seizure-onset zone in non-rapid eye movement (NREM) sleep than in wake or REM sleep. SIGNIFICANCE: The change in spike rates surrounding sleep and seizures differs between temporal and extra-temporal lobe epilepsy. Spikes are more frequent and better localize the SOZ in sleep, particularly in NREM sleep. Quantitative analysis of spikes may provide useful ancillary data to localize the SOZ and improve surgical planning.

Characterizing the neurological phenotype of the hyperinsulinism hyperammonemia syndrome.

BACKGROUND: Hyperinsulinism hyperammonemia (HI/HA) syndrome is caused by activating mutations in GLUD1, encoding glutamate dehydrogenase (GDH). Atypical absence seizures and neuropsychological disorders occur at high rates in this form of hyperinsulinism. Dysregulated central nervous system (CNS) glutamate balance, due to GDH overactivity in the brain, has been hypothesized to play a role. This study aimed to describe the neurologic phenotype in HI/HA syndrome and investigate CNS glutamate levels using glutamate weighted chemical exchange saturation transfer magnetic resonance imaging (GluCEST MRI). In this cross-sectional study, 12 subjects with HI/HA syndrome had plasma ammonia measurement, self- or parent-completed neurocognitive assessments, electroencephalogram (EEG), and GluCEST MRI at 7 T performed. GluCEST MRI measures were compared to a historic reference population of 10 healthy adults. RESULTS: Subjects were five males and seven females with median age of 25.5 years. Seventy-five percent of subjects reported a history of neurodevelopmental problems and 42% had neurocognitive assessment scores outside the normal range. Fifty percent had interictal EEG findings of generalized, irregular spike and wave discharges. Higher variability in hippocampal GluCEST asymmetry (p = 0.002), and in peak hippocampal GluCEST values (p = 0.008), was observed in HI/HA subjects (n = 9 with interpretable MRI) compared to the healthy reference population (n = 10). CONCLUSIONS: The high prevalence of abnormal neurocognitive assessment scores and interictal EEG findings observed highlights the importance of longitudinal neuropsychological assessment for individuals with HI/HA syndrome. Our findings demonstrate the potential application of GluCEST to investigate persistent knowledge gaps in the mechanisms underlying the unique neurophenotype of this disorder.

A framework For brain atlases: Lessons from seizure dynamics.

Brain maps, or atlases, are essential tools for studying brain function and organization. The abundance of available atlases used across the neuroscience literature, however, creates an implicit challenge that may alter the hypotheses and predictions we make about neurological function and pathophysiology. Here, we demonstrate how parcellation scale, shape, anatomical coverage, and other atlas features may impact our prediction of the brain's function from its underlying structure. We show how network topology, structure-function correlation (SFC), and the power to test specific hypotheses about epilepsy pathophysiology may change as a result of atlas choice and atlas features. Through the lens of our disease system, we propose a general framework and algorithm for atlas selection. This framework aims to maximize the descriptive, explanatory, and predictive validity of an atlas. Broadly, our framework strives to provide empirical guidance to neuroscience research utilizing the various atlases published over the last century.

Neurophysiological Evidence for Cognitive Map Formation during Sequence Learning.

Humans deftly parse statistics from sequences. Some theories posit that humans learn these statistics by forming cognitive maps, or underlying representations of the latent space which links items in the sequence. Here, an item in the sequence is a node, and the probability of transitioning between two items is an edge. Sequences can then be generated from walks through the latent space, with different spaces giving rise to different sequence statistics. Individual or group differences in sequence learning can be modeled by changing the time scale over which estimates of transition probabilities are built, or in other words, by changing the amount of temporal discounting. Latent space models with temporal discounting bear a resemblance to models of navigation through Euclidean spaces. However, few explicit links have been made between predictions from Euclidean spatial navigation and neural activity during human sequence learning. Here, we use a combination of behavioral modeling and intracranial encephalography (iEEG) recordings to investigate how neural activity might support the formation of space-like cognitive maps through temporal discounting during sequence learning. Specifically, we acquire human reaction times from a sequential reaction time task, to which we fit a model that formulates the amount of temporal discounting as a single free parameter. From the parameter, we calculate each individual's estimate of the latent space. We find that neural activity reflects these estimates mostly in the temporal lobe, including areas involved in spatial navigation. Similar to spatial navigation, we find that low-dimensional representations of neural activity allow for easy separation of important features, such as modules, in the latent space. Lastly, we take advantage of the high temporal resolution of iEEG data to determine the time scale on which latent spaces are learned. We find that learning typically happens within the first 500 trials, and is modulated by the underlying latent space and the amount of temporal discounting characteristic of each participant. Ultimately, this work provides important links between behavioral models of sequence learning and neural activity during the same behavior, and contextualizes these results within a broader framework of domain general cognitive maps.

Implanting intracranial electrodes does not affect spikes or network connectivity in nearby or connected brain regions.

To determine the effect of implanting electrodes on electrographic features of nearby and connected brain regions in patients with drug-resistant epilepsy, we analyzed intracranial EEG recordings from 10 patients with drug-resistant epilepsy who underwent implant revision (placement of additional electrodes) during their hospitalization. We performed automated spike detection and measured EEG functional networks. We analyzed the original electrodes that remained in place throughout the full EEG recording, and we measured the change in spike rates and network connectivity in these original electrodes in response to implanting new electrodes. There was no change in overall spike rate pre- to post-implant revision (t(9) = 0.1, p = 0.95). The peri-revision change in the distribution of spike rate and connectivity across electrodes was no greater than chance (Monte Carlo method, spikes: p = 0.40, connectivity: p = 0.42). Electrodes closer to or more functionally connected to the revision site had no greater change in spike rate or connectivity than more distant or less connected electrodes. Changes in electrographic features surrounding electrode implantation are no greater than baseline fluctuations occurring throughout the intracranial recording. These findings argue against an implant effect on spikes or network connectivity in nearby or connected brain regions.

Electrocorticography and stereo EEG provide distinct measures of brain connectivity: implications for network models.

Brain network models derived from graph theory have the potential to guide functional neurosurgery, and to improve rates of post-operative seizure freedom for patients with epilepsy. A barrier to applying these models clinically is that intracranial EEG electrode implantation strategies vary by centre, region and country, from cortical grid & strip electrodes (Electrocorticography), to purely stereotactic depth electrodes (Stereo EEG), to a mixture of both. To determine whether models derived from one type of study are broadly applicable to others, we investigate the differences in brain networks mapped by electrocorticography and stereo EEG in a cohort of patients who underwent surgery for temporal lobe epilepsy and achieved a favourable outcome. We show that networks derived from electrocorticography and stereo EEG define distinct relationships between resected and spared tissue, which may be driven by sampling bias of temporal depth electrodes in patients with predominantly cortical grids. We propose a method of correcting for the effect of internodal distance that is specific to electrode type and explore how additional methods for spatially correcting for sampling bias affect network models. Ultimately, we find that smaller surgical targets tend to have lower connectivity with respect to the surrounding network, challenging notions that abnormal connectivity in the epileptogenic zone is typically high. Our findings suggest that effectively applying computational models to localize epileptic networks requires accounting for the effects of spatial sampling, particularly when analysing both electrocorticography and stereo EEG recordings in the same cohort, and that future network studies of epilepsy surgery should also account for differences in focality between resection and ablation. We propose that these findings are broadly relevant to intracranial EEG network modelling in epilepsy and an important step in translating them clinically into patient care.

2021 update to HIV-TRePS: a highly flexible and accurate system for the prediction of treatment response from incomplete baseline information in different healthcare settings.

OBJECTIVES: With the goal of facilitating the use of HIV-TRePS to optimize therapy in settings with limited healthcare resources, we aimed to develop computational models to predict treatment responses accurately in the absence of commonly used baseline data. METHODS: Twelve sets of random forest models were trained using very large, global datasets to predict either the probability of virological response (classifier models) or the absolute change in viral load in response to a new regimen (absolute models) following virological failure. Two 'standard' models were developed with all baseline variables present and 10 others developed without HIV genotype, time on therapy, CD4 count or any combination of the above. RESULTS: The standard classifier models achieved an AUC of 0.89 in cross-validation and independent testing. Models with missing variables achieved AUC values of 0.78-0.90. The standard absolute models made predictions that correlated significantly with observed changes in viral load with a mean absolute error of 0.65 log10 copies HIV RNA/mL in cross-validation and 0.69 log10 copies HIV RNA/mL in independent testing. Models with missing variables achieved values of 0.65-0.75 log10 copies HIV RNA/mL. All models identified alternative regimens that were predicted to be effective for the vast majority of cases where the new regimen prescribed in the clinic failed. All models were significantly better predictors of treatment response than genotyping with rules-based interpretation. CONCLUSIONS: These latest models that predict treatment responses accurately, even when a number of baseline variables are not available, are a major advance with greatly enhanced potential benefit, particularly in resource-limited settings. The only obstacle to realizing this potential is the willingness of healthcare professions to use the system.

The Examination of Instrumental Activities of Daily Living and Perceived Adulthood in Emerging Adults.

Emerging adults differ in terms of the extent to which they perceive themselves as adults. We examined how the ability to perform activities related to independent living (i.e., instrumental activities of daily living [IADLs]) was associated with perceived adulthood. Data were collected from 236 emerging adults in college. Results suggested that IADL scores were positively related to perceived adulthood and achieved criteria of adulthood even after controlling for race, year in school, age, and sex. Results are discussed in terms of the development and importance of IADLs during emerging adulthood.

Novel Approaches to Teaching Aging and Disability: Active Learning Through Design and Exploration.

Active learning emphasizes student engagement and collaboration instead of more passive learning, which involves primarily listening to lectures in the classroom setting. The benefits of active learning are many with an emphasis on the expansion of higher-order processing and critical thinking skills. Active learning can be found in many best practice approaches in the Medicine, Science, Engineering, and Mathematics (MSTEM) fields. Hack-a-thon and hack events are examples of active learning. These are gaining popularity in research institutes, and specifically in engineering, computer science, business, and healthcare settings. Wikipedia defines hack-a-thon as the blending of the words "hack," referring to exploratory programming, and "marathon," referring to a timed event. This article describes a hack-a-thon approach for active learning in the classroom setting.

Predicting Virological Response to HIV Treatment Over Time: A Tool for Settings With Different Definitions of Virological Response.

OBJECTIVE: Definitions of virological response vary from <50 up to 1000 copies of HIV-RNA/mL. Our previous models estimate the probability of HIV drug combinations reducing the viral load to <50 copies/mL, with no indication of whether higher thresholds of response may be achieved. Here, we describe the development of models that predict absolute viral load over time. METHODS: Two sets of random forest models were developed using 50,270 treatment change episodes from more than 20 countries. The models estimated viral load at different time points following the introduction of a new regimen from variables including baseline viral load, CD4 count, and treatment history. One set also used genotypes in their predictions. Independent data sets were used for evaluation. RESULTS: Both models achieved highly significant correlations between predicted and actual viral load changes (r = 0.67-0.68, mean absolute error of 0.73-0.74 log10 copies/mL). The models produced curves of virological response over time. Using failure definitions of <100, 400, or 1000 copies/mL, but not 50 copies/mL, both models were able to identify alternative regimens they predicted to be effective for the majority of cases where the new regimen prescribed in the clinic failed. CONCLUSIONS: These models could be useful for selecting the optimum combination therapy for patients requiring a change in therapy in settings using any definition of virological response. They also give an idea of the likely response curve over time. Given that genotypes are not required, these models could be a useful addition to the HIV-TRePS system for those in resource-limited settings.

2018 update to the HIV-TRePS system: the development of new computational models to predict HIV treatment outcomes, with or without a genotype, with enhanced usability for low-income settings.

Objectives: Optimizing antiretroviral drug combination on an individual basis can be challenging, particularly in settings with limited access to drugs and genotypic resistance testing. Here we describe our latest computational models to predict treatment responses, with or without a genotype, and compare their predictive accuracy with that of genotyping. Methods: Random forest models were trained to predict the probability of virological response to a new therapy introduced following virological failure using up to 50 000 treatment change episodes (TCEs) without a genotype and 18 000 TCEs including genotypes. Independent data sets were used to evaluate the models. This study tested the effects on model accuracy of relaxing the baseline data timing windows, the use of a new filter to exclude probable non-adherent cases and the addition of maraviroc, tipranavir and elvitegravir to the system. Results: The no-genotype models achieved area under the receiver operator characteristic curve (AUC) values of 0.82 and 0.81 using the standard and relaxed baseline data windows, respectively. The genotype models achieved AUC values of 0.86 with the new non-adherence filter and 0.84 without. Both sets of models were significantly more accurate than genotyping with rules-based interpretation, which achieved AUC values of only 0.55-0.63, and were marginally more accurate than previous models. The models were able to identify alternative regimens that were predicted to be effective for the vast majority of cases in which the new regimen prescribed in the clinic failed. Conclusions: These latest global models predict treatment responses accurately even without a genotype and have the potential to help optimize therapy, particularly in resource-limited settings.

An update to the HIV-TRePS system: the development and evaluation of new global and local computational models to predict HIV treatment outcomes, with or without a genotype.

OBJECTIVES: Optimizing antiretroviral drug combination on an individual basis in resource-limited settings is challenging because of the limited availability of drugs and genotypic resistance testing. Here, we describe our latest computational models to predict treatment responses, with or without a genotype, and compare the potential utility of global and local models as a treatment tool for South Africa. METHODS: Global random forest models were trained to predict the probability of virological response to therapy following virological failure using 29 574 treatment change episodes (TCEs) without a genotype, 3179 of which were from South Africa and were used to develop local models. In addition, 15 130 TCEs including genotypes were used to develop another set of models. The 'no-genotype' models were tested with an independent global test set (n = 1700) plus a subset from South Africa (n = 222). The genotype models were tested with 750 independent cases. RESULTS: The global no-genotype models achieved area under the receiver-operating characteristic curve (AUC) values of 0.82 and 0.79 with the global and South African tests sets, respectively, and the South African models achieved AUCs of 0.70 and 0.79. The genotype models achieved an AUC of 0.84. The global no-genotype models identified more alternative, locally available regimens that were predicted to be effective for cases that failed their new regimen in the South African clinics than the local models. Both sets of models were significantly more accurate predictors of outcomes than genotyping with rules-based interpretation. CONCLUSIONS: These latest global models predict treatment responses accurately even without a genotype, out-performed the local South African models and have the potential to help optimize therapy, particularly in resource-limited settings.

Computational models as predictors of HIV treatment outcomes for the Phidisa cohort in South Africa.

BACKGROUND: Selecting the optimal combination of HIV drugs for an individual in resource-limited settings is challenging because of the limited availability of drugs and genotyping. OBJECTIVE: The evaluation as a potential treatment support tool of computational models that predict response to therapy without a genotype, using cases from the Phidisa cohort in South Africa. METHODS: Cases from Phidisa of treatment change following failure were identified that had the following data available: baseline CD4 count and viral load, details of failing and previous antiretroviral drugs, drugs in new regimen and time to follow-up. The HIV Resistance Response Database Initiative's (RDI's) models used these data to predict the probability of a viral load < 50 copies/mL at follow-up. The models were also used to identify effective alternative combinations of three locally available drugs. RESULTS: The models achieved accuracy (area under the receiver-operator characteristic curve) of 0.72 when predicting response to therapy, which is less accurate than for an independent global test set (0.80) but at least comparable to that of genotyping with rules-based interpretation. The models were able to identify alternative locally available three-drug regimens that were predicted to be effective in 69% of all cases and 62% of those whose new treatment failed in the clinic. CONCLUSION: The predictive accuracy of the models for these South African patients together with the results of previous studies suggest that the RDI's models have the potential to optimise treatment selection and reduce virological failure in different patient populations, without the use of a genotype.

Computational Models as Predictors of HIV Treatment Outcomes for the Phidisa Cohort in South Africa

Background: Selecting the optimal combination of HIV drugs for an individual in resourcelimited settings is challenging because of the limited availability of drugs and genotyping.Objective: The evaluation as a potential treatment support tool of computational models that predict response to therapy without a genotype; using cases from the Phidisa cohort in South Africa.Methods: Cases from Phidisa of treatment change following failure were identified that had the following data available: baseline CD4 count and viral load; details of failing and previous antiretroviral drugs; drugs in new regimen and time to follow-up. The HIV Resistance Response Database Initiative's (RDI's) models used these data to predict the probability of a viral load 50 copies/mL at follow-up. The models were also used to identify effective alternative combinations of three locally available drugs.Results: The models achieved accuracy (area under the receiver-operator characteristic curve) of 0.72 when predicting response to therapy; which is less accurate than for an independent global test set (0.80) but at least comparable to that of genotyping with rules-based interpretation. The models were able to identify alternative locally available three-drug regimens that were predicted to be effective in 69% of all cases and 62% of those whose new treatment failed in the clinic.Conclusion: The predictive accuracy of the models for these South African patients together with the results of previous studies suggest that the RDI's models have the potential to optimise treatment selection and reduce virological failure in different patient populations; without the use of a genotype