Evidence Based MedicineMRCEM Part A

Dr Sajjad Pathan

Sensitivity & Specificity

Test Positive Test Negative

Disease TP FN Sn = TP / (TP +FN)

No Disease FP TN Sp = TN / (TN + FP)

PPV = TP / (TP + FP)

NPV = TN/ (TN+FN)

Sensitivity rules out disease (SNOUT)

Specificity Rules in disease (SPIN)

PPV is probability of having a disease if the test is positive

NPV is probability of not having a disease if the test is negative

PVs are affected by Prevalence

Trues are always on TOP

Likelihood Ratios Likelihood Ratios are very great tool for assessing diagnostic accuracy of a

symptom or sign or tests

Positive Likelihood Ratio (LR+) = Sn / (1 – Sp)

Negative Likelihood Ratio (LR-) = Sp / (1 – Sn)

Likelihood Ratio >1 is showing test is associated with diseases and <1 is showing test is not associated with diseases

Likelihood ratio of 1 is useless

However, LR+ >5 and LR- < 0.2 are very useful

Incidence & Prevalence Number of New Cases / Population at risk = Incidence Total Number of Cases/ Population at Risk = Prevalence Prevalence = Incidence x Duration (P = ID) ~ PID If duration of Illness is short, Prevalence ~ Incidence Similarly, If Duration of Illness is Longer (Chronic Conditions),

Prevalence > Incidence Note: Prevalence affects Predictive Value If Prevalence Increases, PPV Increases and NPV Decreases Sn & Sp are not affected by Prevalence

Types of Study Retrospective Studies: Will see control group and case group and see if the

risk factor was present. ODD Ratio = Case: Control

Prospective: Cohort study. Relative Risk

Cross Sectional Study: Looks at disease prevalence

ProspectiveRetrospective

Simple Normal Distribution Curve ( Gaussian Curve)

Mean = Algebraic average

Mode = Maximum appearing variable

Median = Middle Term

In a Simple Normal Bell Shaped Curve: Mean = Mode = Median = Peak of Curve

Data is equally distributed (symmetrical)

1 SD each side is 34 % (Total both sides 68 %)

2 SD = 95 %

3 SD = 99.7%

Normal Distribution Curve

Absolute Risk Reduction Consider we have 2 groups of patients with pain, we divide them equally in

Control group and Experimental Group.

Control Group Event Rate (CER) = 30/100 (30%)

Experimental Group Event Rate (EER) = 80/100 (80%)

ARR = EER – CER = 50% = 0.5

NNT = 1/ARR = 1/0.5 = 2

Drug Given Benefit + No Benefit

Experimental Drug 80 20

Placebo Drug 30 70

Numbers Needed to Treat Number of patients needed to treat to benefit 1 patient (prevent 1 bad

outcome)

Ideal or perfect NNT is 1 (5 – 20 is much more than useful and still represents good therapeutic benefit)

NNT = 1/ARR

Aspiring in AMI has NNT of 40

NNT should be as low as possible but not “0”, and Numbers needed to harm should be as high as possible

Types of Data Can be Qualitative or Quantitative

Nominal Ordinal Interval Ratios

NOn Parametric (Nominal, Ordinal)

Parametric

Central Measure

Median (Skewed Data) Mean (Normally Distributed Data)

Correlation Test

Spearman Pearson

Other Tests Mann Whitney, Wilcoxon, Friedmann

T Test, ANOVA

Types of BiasSelection Bias Randomization

Recall Bias

Measurement Bias

Procedure Bias

Observer Expectancy Bias

Cofounding Bias

Lead Time Bias

P Value & Types of Errors Whenever you see a study, they run stats and say the P Value (alpha) was

found to be < 0.05 and therefore results are significant.

What do you understand by that?

What is a hypothesis?

What is alpha error or type 1 error

What is beta error or type 2 error

Which one is BAD?

What you can do about it?

Hypothesis is any statement which we want to prove

Ex PCOD Patients are prone to be diabetic

Then we make an opposite hypothesis. We call it null hypothesis

We collect data and analyze it using one of the test.

The test says that, P Value is less than 0.05, which means that the null hypothesis is false, that is there is less than 5 % chance that the things has happened by chance

Type 1 error (alpha error) is rejecting Null hypothesis when it is true

Type 2 error (Beta error) is accepting Null hypothesis when it is actually not true

Type 1 is worse than type 2

We reduce errors by increasing sample size (More people = More POWER)

Power = 1 - beta

Blinding in Clinical Trials In Double Blind Study, Neither the investigator nor the subjects are aware of

which intervention they are being given. Ideally those analyzing data are also unaware of the interventions, but it is not possible always. Those interpreting data may know the intervention without affecting the results.

Blinding helps to reduce BIASES

Publication Biases are bias by the researchers, pharma and journals that are biased to produce positive results