Search PubMed⌕ Search

SEARCH · Search PubMed

Results for “Explainability”

Search indexed PubMed citations on genomics, clinical trials, systematic reviews and public health. Explore titles, authors and supplied subject terms, then open the PubMed record.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 19 recordsLinked to original sources
PubMed1991

HF-Explain: a natural language generation system for explaining a medical expert system.

Causal models have been used, with considerable success, to reason in the medical domain. While these systems typically have a robust reasoning mechanism and knowledge base about their specific area of expertise, their ability to satisfactorily explain their results in a meaningful, coherent and concise manner has been less impressive then their diagnostic capabilities. This paper describes a program, HF-Explain, that generates natural language explanations of one such system--the Heart Failure Program. HF-Explain, is loosely based on work done by McKeown in the Text system, using augmented transition networks (ATN) as a formalism to guide the explanation process. The result is a coherent, concise, accurate and rich explanation of Heart Failure Programs' diagnostic hypotheses.

Cardiac Output, Low↗
PubMed2003

Explaining variations in reported diabetes prevalence in general practice: how much variation is explained by differences between practice populations?

There are large variations in reported diabetes prevalence within United Kingdom (UK) populations. Linear regression was used to investigate whether population characteristics could explain the variation in prevalence between 19 practices with relatively complete diagnostic recording. Population obesity and South Asian ethnicity largely explained observed variation in prevalence (adjusted R2 = 0.80). When adjusted for obesity and ethnicity, the deprivation score was no longer a predictor of diabetes prevalence. If true variation in prevalence between practices is largely predicted by population ethnicity and obesity, these population characteristics could be used to predict expected prevalence and to assess the completeness of practice registers.

Aged↗
PubMed1989

On approaches to explaining cardiovascular reactivity: toward explanations that explain.

The control-system mode of analysis proposed by Pavloski for cardiovascular reactivity phenomena is a purposive or teleological form of explanation, because the explicans has purpose as the central term. The problem with such explanations is that they are circular and empty, providing only a rhetorical feeling of understanding, and absence of strong-inference research. The S-O-R approach, which Pavlovski opposes, provides at least the potential for non-circular explanations, provided that the O-related explanatory constructs are specified in normal cause-effect terms so that, in another context, the explicans can also serve as an explicandum, and vice-versa. Pavloski may be right in suggesting that the evidence requires a feedback, "control-system" construct to account for cardiovascular reactivity phenomena (although he is far from having proved this in his paper), but such a view would need to be supplemented by (non-purposive) control-system explanations that really explain in the logical rather than the merely rhetorical sense of that term.

Arousal↗
PubMed1997

Variation in prostate cancer survival explained by significant prognostic factors.

PURPOSE: Traditional survival analytical tools do not reveal the ability of significant prognostic factors to predict (that is, explain variation in) survival. We used survival data in patients with prostate cancer to illustrate how the association of factors with survival diverges from their ability to explain variation in survival; bladder cancer was included as a point of general comparison. MATERIALS AND METHODS: We used the 1973 to 1990 records of the Surveillance Epidemiology and End Results program. Multivariate proportional hazards models were used to identify factors that significantly associated with survival. The proportion of variation explained by these factors was estimated with the Schemper method. RESULTS: The dataset included 10,636 patients with prostate cancer and 1,070 with bladder cancer. Median survival was significantly longer in prostate than bladder cancer; other characteristics were similarly distributed. Age, stage and marital status were associated with survival in both cancers (p value range 0.0001 to 0.0009). The total proportion of variation explained was 7.1% and 32.1% for prostate and bladder cancer, respectively. In prostate cancer, age, stage and marital status explained 0.6, 5.5 and 0.4%, of the adjusted proportion of variation explained, respectively, and in bladder cancer, they explained 14.7, 8.9 and 0.6%, respectively. CONCLUSIONS: Proportional hazards models identified but did not reveal the ability of significant prognostic factors to explain variations in survival. The proportion of variation explained analyses illustrate why predicting survival is so difficult, especially in prostate cancer. The prognostic factors used do not possess the ability to explain variation in survival; new prognostic factors must be identified.

Aged↗
PubMed2005

Explained variation in a model of therapeutic decision making is partitioned across patient, physician, and clinic factors.

BACKGROUND AND OBJECTIVE: Data on therapeutic decision making have a multilevel structure that can include patient-, provider-, and facility-level variables. A statistical method is presented for attributing explained variation in patient care to different levels of aggregation in a multilevel model with the aim of prioritizing and targeting quality improvement interventions. STUDY DESIGN AND SETTING: The proposed method is used in an analysis of adherence to evidence-based guidelines for the care of patients at risk of osteoporosis. Explained variation from a multilevel model of appropriate care is partitioned across patient-, physician-, and clinic-level factors. RESULTS: The combination of patient, physician, and clinic factors explained 20.0% of the variation in patient care. Individual physician effects explained 14.0% of the variation in the data; however, more than half of this explained variation could have been attributed to the individual clinic effect. Patient fixed effects alone explained 13.4% of the variation in the observed clinical decisions. CONCLUSION: The proposed approach is an intuitive and statistically valid method for attributing explained variation in a multilevel analysis of therapeutic decision making.

Administration, Oral↗
PubMed2004

Haplotypes and SNPs in 13 lipid-relevant genes explain most of the genetic variance in high-density lipoprotein and low-density lipoprotein cholesterol.

Single nucleotide polymorphisms (SNPs) and derived haplotypes within multiple genes may explain genetic variance in complex traits; however, this hypothesis has not been rigorously tested. In an earlier study we analyzed six genes and have now expanded this investigation to include 13. We studied 250 families including 1054 individuals and measured lipid phenotypes. We focused on low-density cholesterol (LDL), high-density cholesterol (HDL) and their ratio (LDL/HDL). A component analysis of the phenotypic variance relying on a standard genetic model' showed that the genetic variance on LDL explained 26%, on HDL explained 38% and on LDL/HDL explained 28% of the total variance, respectively. Genotyping of 93 SNPs in 13 lipid-relevant genes generated 230 haplotypes. The association of haplotypes in all the genes tested explained a major fraction of the genetic phenotypic variance component. For LDL, the association with haplotypes explained 67% and for HDL 58% of the genetic variance relative to the polygenic background. We conclude that these haplotypes explain most of the genetic variance in LDL, HDL and LDL/HDL in these representative German families. An analysis of the contribution to the genetic variance at each locus showed that APOE (50%), CETP (28%), LIPC (9%), APOB (8%) and LDLR (5%) influenced variation in LDL. LIPC (53%), CETP (25%), ABCA1 (10%), LPL (6%) and LDLR (6%) influenced the HDL variance. The LDL/HDL ratio was primarily influenced by APOE (36%), CETP (27%) and LIPC (31%). This expanded analysis substantially increases the explanation of genetic variance on these complex traits.

Cholesterol, HDL↗
PubMed2006

The development of a Local Index of Need (LIN) and its use to explain variations in social services expenditure on mental health care in England.

Wide variations in deprivation exist across England and it is likely that these variations translate into differences in the need for mental health care. Social Services Departments in England account for a substantial proportion of mental health expenditure. It is important that the distribution of expenditure reflects mental health needs. This paper's aims are to (1) describe the development of a new indicator of mental health needs, (2) use the index to explain variations in social services expenditure on mental health, and (3) compare the index with other established measures of need. A principal components analysis of sociodemographic variables considered to be indicators of need was used to produce four distinct factors for 148 Local Authority areas in England. A weighted sum of these factors was used to produce a single index. (Weights were the proportion of variance explained by each factor.) The index was used in a regression model to explain variations in spending on mental health care and was compared with (1) a model containing the four individual factors, (2) the current method of allocating resources, (3) the index used to allocate resources to primary care trusts, (4) the Mental Illness Needs Index (MINI), (5) four indices of deprivation produced by the Office of the Deputy Prime Minister, and (6) the average of the above four indices. The new index could explain 54% of variation, compared with 56% using the current method. The four-factor model could explain 66%, whilst the other models could explain between 37% and 20%. This new index has the advantage that it is not based on previous levels of utilisation or expenditure and yet still explains a comparable amount of variation as the current method. However, a disaggregated model containing individual factors may be preferable.

Adolescent↗
PubMed2006

Do risk factors differ between explained sudden unexpected death in infancy and sudden infant death syndrome?

BACKGROUND: In Germany, 2910 infants died in 2004; for many infants the reason was clear, especially prematurity or congenital abnormalities. However, 394 babies die every year suddenly and unexpectedly. The cause may be immediately clear, but is often not obvious. AIMS: (1) To describe the causes of explained sudden unexpected death in infancy (SUDI) and (2) to compare risk factors for sudden infant death syndrome (SIDS) and explained SUDI. METHODS: A 3-year population-based case-control study in Germany, 1998-2001. RESULTS: 455 deaths, of which 51 (11.2%) were explained. Most of these deaths were due to respiratory or generalised infections. The risk factors for SIDS and explained SUDI were remarkably similar except for sleep position and breast feeding. Prone sleeping position is a major risk factor for SIDS (adjusted odds ratio (OR) 7.16, 95% confidence interval (CI) 3.85 to 13.31) but not for explained SUDI (adjusted OR 1.71, 95% CI 0.25 to 11.57). Not being breast fed in the first 2 weeks of life is a risk factor for SIDS (adjusted OR 2.37, 95% CI 1.46 to 3.84) but not for explained SUDI (adjusted OR 0.39, 95% CI 0.08 to 1.83). CONCLUSIONS: Prone sleeping position is a unique risk factor for SIDS. Socioeconomic disadvantage and maternal smoking are risk factors for both SIDS and explained SUDI, and provide an opportunity for targeted intervention.

Age Distribution↗
PubMed2001

The role of income differences in explaining social inequalities in self rated health in Sweden and Britain.

STUDY OBJECTIVE: To analyse to what extent differences in income, using two distinct measures-as distribution across quintiles and poverty-explain social inequalities in self rated health, for men and women, in Sweden and Britain. DESIGN: Series of cross sectional surveys, the Swedish Survey of Living Conditions (ULF) and the British General Household Survey (GHS), during the period 1992-95. PARTICIPANTS AND SETTING: Swedish and British men and women aged 25-64 years. Approximately 4000 Swedes and 12 500 Britons are interviewed each year in the cross sectional studies used. The sample contains 15 766 people in the Swedish dataset and 49 604 people in the British dataset. MAIN RESULTS: The magnitude of social inequalities in less than good self rated health was similar in Sweden and in Britain, but adjusting for income differences explained a greater part of these in Britain than in Sweden. In Britain the distribution across income quintiles explained 47% of the social inequalities in self rated health among women and 31% among men, while in Sweden it explained, for women 13% and for men 20%. Poverty explained 22% for British women and 8% for British men of the social inequalities in self rated health, while in Sweden poverty explained much less (men 2.5% and women 0%). CONCLUSIONS: The magnitude of social inequalities in self rated health was similar in Sweden and in Britain. However, the distribution of income across occupational social classes explains a larger part of these inequalities in Britain than in Sweden. One reason for this may be the differential exposure to low income and poverty in the two countries.

Adult↗
PubMed1999

How much disability in rheumatoid arthritis is explained by rheumatoid arthritis?

OBJECTIVE: To measure the proportion of disability explained by disease manifestations compared with nondisease factors in rheumatoid arthritis (RA). METHODS: A hypothetical model of the disablement process specific for RA was constructed using the demographic, sociocultural, and clinical characteristics of a consecutive cohort of RA patients. Disability was measured with the modified Health Assessment Questionnaire (M-HAQ) and the physical function scale of the Medical Outcomes Study Short Form 36 (SF-36) questionnaire. Independent variables, grouped according to their position in the RA disablement process model, were sequentially entered in a series of hierarchical regression models. The proportion of variance in disability explained by each group of variables was measured by the group's incremental R2. RESULTS: The overall proportion of disability explained by the full model was 59%. Factors in the main disease-disability pathway explained 33%, of which 3% was explained by disease duration, 5% by the Westergren erythrocyte sedimentation rate, 14% by articular signs and symptoms, and 11% by performance-based functional limitations. External modifiers and contextual variables explained 26% of the variance in disability, of which age and sex accounted for 2%, formal education 4%, psychological status 17%, and symptoms of depression 3%. CONCLUSION: Both the main disease-disability pathway and factors external to this pathway contribute significantly to disability in RA. These findings provide evidence of the relative influence of psychosocial factors, compared with disease manifestations, on the disability of patients with RA.

Adult↗
PubMed2001

Testing Stereotype Threat: Does Anxiety Explain Race and Sex Differences in Achievement?

Steele's (1992, 1997) stereotype-threat theory attempts to explain underperformance of minority students in academic domains and of women in mathematics. Steele argues that situational self-relevance of negative group stereotypes in testing situations increases the anxiety these students experience and that these differential anxiety levels explain performance differences. Research shows that manipulation of stereotype threat can affect academic performance. However, there has been little research testing whether anxiety does at least partially explain the relationship between race and achievement. The goal of this study was to examine whether anxiety will explain racial differences in academic performance and gender differences in math performance in the context of a nationally representative sample of high school seniors. Partial mediation was observed, with anxiety explaining significant portions of the racial differences in academic performance. Anxiety also partially explained sex differences in math achievement, although the effect sizes were very small. These results provide general support for Steele's stereotype-threat hypothesis. Copyright 2001 Academic Press.

Journal Article↗
PubMed2003

How much variance is explained by ecologists? Additional perspectives.

A recent meta-analysis of meta-analyses by Møller and Jennions suggested that ecologists using statistical models are explaining between 2.5% and 5.42% of the variability in ecological studies. Although we agree that there is considerable variability in ecological systems that is not explained, we disagree with the approach and general conclusions of Møller and Jennions. As an alternate perspective, we explored the question: "How much ecological variation in relationships is not explained?" We did this by examining published studies in five different journals representative of the numerous sub-disciplines of ecology. We quantified the proportion of variance not explained in statistical models as the residual or random error compared to the total variation in the data set. Our results indicate that statistical models explain roughly half of the variation in variables of interest, vastly different from the 2.5%-5.42% reported by Møller and Jennions. This difference resulted largely from a different level of analysis: we considered the original study to be the appropriate level for quantifying variability while Møller and Jennions combined studies at different temporal and spatial scales and attempted to find universal single-factor relationships between ecological variables across study organisms or locations. Therefore, we believe that Møller and Jennions actually measured the universality of single factor effects across multiple ecological systems, not the amount of variability in ecological studies explained by ecologists. This study, combined with Møller and Jennions', illustrates importance of applying statistical models appropriately to assess ecological relationships.

Animals↗
PubMed1988

A model of force production that explains the lag between crossbridge attachment and force after electrical stimulation of striated muscle fibers.

Whereas the mechanical behavior of fully activated fibers can be explained by assuming that attached force-producing crossbridges exist in at least two configurations, one exerting more force than the other (Huxley A. F., and R. M. Simmons. 1971. Nature [Lond.]. 233:533-538), and the behavior of relaxed fibers can be explained by assuming a single population of weakly binding rapid-equilibrium crossbridges (Schoenberg, M. 1988. Biophys. J. 54:135-148), it has not been possible to explain the transition between rest and activation in these terms. The difficulty in explaining why, after electrical stimulation of resting intact frog skeletal muscle fibers at 1-5 degrees C, force development lags stiffness development by more than 15 ms has led a number of investigators to postulate additional crossbridge states. However, postulation of an additional crossbridge state will not explain the following three observations: (a) Although the lag between force and stiffness is very different after stimulation, during the redevelopment of force after an extended period of high velocity shortening, and during relaxation of a tetanus, nonetheless, the plots of force versus stiffness in each of these cases are approximately the same. (b) When the lag between stiffness and force during the rising phase of a twitch is changed nearly fourfold by changing temperature, again the plot of force versus stiffness remains essentially unchanged. (c) When a muscle fiber is subjected to a small quick length change, the rate constant for the isometric force recovery is faster when the length change is applied during the rising phase of a tenanus than when it is applied on the plateau. We have been able to explain all the above findings using a model for force production that is similar to the 1971 model of Huxley and Simmons, but which makes the additional assumption that the force-producing transition envisioned by them is a cooperative one, with the back rate constant of the force-producing transition decreasing as more crossbridges attach.

Animals↗
PubMed1998

The conflict theory of genomic imprinting: how much can be explained?

In some mammalian genes, paternally and maternally derived alleles are expressed differently: this phenomenon is called genomic imprinting. Several-explanations have been proposed for the observed patterns of genomic imprinting, but the most successful explanation is the genetic conflict hypothesis--natural selection operating on the gene expression produces the parental origin-dependent gene expression--because the paternally derived allele tends to be less related to the siblings of the same mother than the maternal allele and hence the paternal allele should evolve to be more aggressive in obtaining maternal resources. The successes and failures of this argument have been examined in explaining the observed patterns of genomic imprinting in mammals. After a brief summary of the observations with some examples, a quantitative genetic model describing the evolution of the cis-regulating element of a gene affecting the maternal resource acquisition was presented. The model supports the verbal argument that the growth enhancer should evolve to show imprinting with the paternal allele expressed and the maternal allele inactive, whereas a growth suppressor gene tends to have an inactive paternal allele and an active maternal allele. There are four major problems of the genetic conflict hypothesis. (1) Some genes affect embryonic growth but are not imprinted (e.g., Igf1), which can be explained by considering recessive, deleterious mutations on the coding regions, (2) A gene exists that shows the pattern that is a perfect reversal (Mash2), which is needed for placental growth, and yet has an active maternal allele and an inactive paternal allele. This can be explained if the overproduction of this gene causes dose-sensitive abortion to occur in early gestation. (3) Paternal disomies are sometimes smaller than normal embryos. This is a likely outcome of evolution if imprinted genes control the allocation between placenta and embryo by modifying the cell developmental fate. (4) Genes on X chromosomes do not follow the predictions of the genetic conflict hypothesis. For genes on X chromosomes, two additional forces of natural selection (sex differentiation and dosage compensation) cause genomic imprinting, possibly in the opposite direction. Available evidence suggests that these processes are stronger than the natural selection caused by female multiple mating. Finally, the same formalism of evolution can handle an alternative nonconflict hypothesis: genomic imprinting might have evolved because it reduces the risk of the spontaneous development of parthenogenetic embryo, causing a serious threat to the life of the mother (ovarian time bomb hypothesis). This hypothesis can also explain major patterns of genomic imprinting. In conclusion, the genetic conflict hypothesis is very successful in explaining the observed patterns of imprinting for autosomal genes and probably is the most likely evolutionary explanation for them. However, for genes on X chromosomes, other processes of natural selection are more important. Considering that a nonconflict hypothesis can also explain the patterns in principle, we need a quantitative estimate of various parameters, such as the rate of dose-dependent abortion, the degree of female promiscuity, and the rate of spontaneous development of the parthenogenetic embryo, in order to make judgments on the relative importance of different forces of natural selection to form genomic imprinting.

Abortion, Veterinary↗
PubMed2000

A clinical comparison of SIDS and explained sudden infant deaths: how healthy and how normal? CESDI SUDI Research Group. Confidential Inquiry into Stillbirths and Deaths in Infancy study.

OBJECTIVES: To compare the clinical characteristics associated with sudden infant death syndrome (SIDS) and explained sudden unexpected deaths in infancy (SUDI). DESIGN: Three year population based, case control study with parental interviews for each death and four age matched controls. SETTING: Five regions in England (population, > 17 million; live births, > 470,000). SUBJECTS: SIDS: 325 infants; explained SUDI: 72 infants; controls: 1,588 infants. RESULTS: In the univariate analysis, all the clinical features and health markers at birth, after discharge from hospital, during life, and shortly before death, significant among the infants with SIDS were in the same direction among the infants who died of explained SUDI. In the multivariate analysis, at least one apparent life threatening event had been experienced by more of the infants who died than in controls (SIDS: 12% v 3% controls; odds ratio (OR) = 2.55; 95% confidence interval (CI), 1.02 to 6.41; explained SUDI: 15% v 4% controls; OR = 16.81; 95% CI, 2.52 to 112.30). Using a retrospective illness scoring system based on "Baby Check", both index groups showed significant markers of illness in the last 24 hours (SIDS: 22% v 8% controls; OR = 4.17; 95% CI, 1.88 to 9.24; explained SUDI: 49% v 8% controls; OR = 31.20; 95% CI, 6.93 to 140.5). CONCLUSIONS: The clinical characteristics of SIDS and explained SUDI are similar. Baby Check might help identify seriously ill babies at risk of sudden death, particularly in high risk infants.

Case-Control Studies↗
PubMed1999

Epidemiology of SIDS and explained sudden infant deaths. CESDI SUDI Research Group.

OBJECTIVES: To establish whether epidemiologic characteristics for sudden infant death syndrome (SIDS) have changed since the decrease in death rate after the "Back to Sleep" campaign in 1991, and to compare these characteristics with sudden and unexpected deaths in infancy (SUDI) from explained causes. DESIGN: Three-year, population-based, case-control study. Parental interviews were conducted soon after the death and for 4 controls matched for age and date of interview. All sudden unexpected deaths were included in the study and the cause of death was established by a multidisciplinary panel of the relevant health care professionals taking into account past medical and social history of the mother and infant, the circumstances of death, and a full pediatric postmortem examination. Contributory factors and the final classification of death were made using the Avon clinicopathologic system. SETTING: Five regions in England, with a total population of >17 million people, took part in the study. The number of live births within these regions during the particular time each region was involved in the study was 473 000. STUDY PARTICIPANTS: Three hundred twenty-five SIDS infants (91.3% of those available), 72 explained SUDI infants (86.7% of those available), and 1588 matched control infants (100% of total for cases included). RESULTS: Many of the epidemiologic features that characterize SIDS infants and families have remained the same, despite the recent decrease in SIDS incidence in the United Kingdom. These include the same characteristic age distribution, few deaths in the first few weeks of life or after 6 months, with a peak between 4 and 16 weeks, a higher incidence in males, lower birth weight, shorter gestation, and more neonatal problems at delivery. As in previous studies there was a strong correlation with young maternal age and higher parity and the risk increased for infants of single mothers and for multiple births. A small but significant proportion of index mothers had also experienced a previous stillbirth or infant death. The majority of the SIDS deaths (83%) occurred during the night sleep and there was no particular day of the week on which a significantly higher proportion of deaths occurred. Major epidemiologic features to change since the decrease in SIDS rate include a reduction in the previous high winter peaks of death and a shift of SIDS families to the more deprived social grouping. Just more than one quarter of the SIDS deaths (27%) occurred in the 3 winter months (December through February) in the 3 years of this study. In half of the SIDS families (49%), the lone parent or both parents were unemployed compared with less than a fifth of control families (18%). This difference was not explained by an excess of single mothers in the index group. Many of the significant factors relating to the SIDS infants and families that distinguish them from the normal population did not distinguish between SIDS and explained SUDI. In the univariate analysis many of the epidemiologic characteristics significant among the SIDS group were also identified and in the same direction among the infants dying as SUDI attributable to known causes. The explained deaths were similarly characterized by the same infant, maternal, and social factors, 48% of these families received no waged income. Using logistic regression to make a direct comparison between the two index groups there were only three significant differences between the two groups of deaths: 1) a different age distribution, the age distribution of the explained deaths peaked in the first 2 months and was more uniform thereafter; 2) more congenital anomalies were noted at birth (odds ratio [OR] = 3.14; 95% confidence intervals [CI]: 1.52-6. (ABSTRACT TRUNCATED)

Age Distribution↗