|
CLASSIF1 Individual Patient Classification:
SAIS1 Trial (Sitagliptin/Glimepiride)
Günter Valet
|
< back
=external link
Summary: HbA1c (glycated hemoglobin A1c) is the best
discriminating of the
32
available SAIS1
clinical trial parameters.
It identifies Sitagliptin and Glimerpiride treated
patients with specificities of 97.4% and 95.3% at
sensitivities of 62.9% and 76.3% for the training set patients
with 100.0% and 88.3% specificities at 75.0% and 63.6%
sensitvities for the validation patients, using the
5/95% percentile thresholds for
CLASSIF1
algorithmic classification.
The less discriminating but parameter fishing percentile thresholds
25/75% and 30/70% identify patients
(tab.2)
with 85.0% and 70.7%% positive predictive
values in the training and 80.0% and 77.8% in the validation
set.
Glimepiride therapy induces the single decrease of the
HbA1c parameter while Sitagliptin as
DPP-4 (dipeptidyl peptidase 4) inhibitor is accompanied after 26 weeks of
therapy by
increases of
(+) CPEPTI (insulin C-peptide index),
(+) HDL-cholesterin (high density lipoprotein-cholesterin) classification mask parameters
in combination with parameters of the oxidative metabolism (+) SOD
(superoxide dismutase) and (+) BAP (biological antioxidant potential),
although when compared to Glimepiride after therapy,
reduced
oxidative stimulation (-) D-ROMS (reactive oxygen metabolites-derived
compounds) at increased proinflammatory parameters (+)
CRP (C-reactive protein), (+) TNF-alpha (tumor necrosis factor alpha
preexisting !),
as well as (+) TPRI (peripheral vascular resistance) at decreased
(-) CARDIX (cardiac output index) are observed for Sitagliptin.
New patients can be
manually
classified without computer as an advantage of the CLASSIF1 algorithm
for everyday medical.practice, in case classification masks and
relevant percentile thresholds are publicly available.
1. Introduction
Clinical trial data are usually statistically evaluated
(1
2)
or more recently by machine learning algorithms
(3).
Since ultimately individual patients are treated, the utility of
individual patient oriented software like the CLASSIF1 algorithm
(4,
5)
was investigated using the SAIS1 EXCEL data of
(weeks
0
and
26)
2. Sitagliptin after Therapy (26 weeks)
fig.1 Data classification with the 5/95% percentile threshold pair
shows the high accuracy classification potential of
HbA1c
(glycated hemoglobin A1c),
provided additional parameters of similar selectivity can be found for
reliable (>95% or >99% accuracy) clinical classifications.
(training
as well as
validation patients and
parameter lists are attached as pdf files).
Average recognition
(ARV=(specificity+sesitivity)/2 (%)) and
average predictive values (APV=(npv+ppv)/2 (%)) were 80.2 and
84.8 %.
Validation patients classify similarly with ARV and APV values of 87.5
and 91.5%.
fig.2 Change of the percentile pair from 5/95% to the less stringent
condition of 25/75% results in a less accurate classification but besides
HbA1c points with four additionally selected parameters in the
classification mask to potential molecular mechanisms involved in Sitagliptin
action.
(training and
validation lists,
classification,
means with
significances, percentile thresholds).
ARV and APV of 70.0 and 74.6 % at the 25/75%
percentile thresholds are lower than for the 5/95% classification
(fig.1),
demonstrating increased selectivity of the 5/95% thresholds.for accurate
classifications.
The validation patients classify with APV=70.0% and APV=74,6%
equally insufficiently.
fig.3 The parameters selected in the classification masks concern
lipid and oxidatative pathways besides the high discrimination
HbA1c parameter
(fig.1).
The mask parameter sequence
(-+++++) (-)
HBA1C
(hemoglobin A1c), (+)
CPEPTIND
(C-peptide index ), (+) HDL (high density lipoprotein cholesterol),
(+) SOD (superoxide dismutase), (+)
BAP
(biological antioxidant potential)
(p<0.05, <
0.10).
represents the location of the respective
data columns in the database
Data columns nr.13,18,21,25,32 with
(means, statistics, percentiles)
were selected from initially 32 data columns, that is
most of the available information does not contribute to the
classificarion result.
.
3. Glimepiride after Therapy (26 weeks)
fig.4 Glimepiride classifies with the 5/95% percentiles results in
an ARV of 80.6 and an APV of 84.3%, similarly
as for the Sitagliptin patients
(fig.1).
fig.5 HbA1c in Glimepiride as in Sitagliptin treated patients
is the only well discriminating parameter and shows high statistical
significance.The HbA1c value distribution at 0 weels is right skewed
as evidenced by the lower median of 7.29% as compared to the
mean of 7.45%
This provides good discrimination at the left flank of the value distribution
curve, explaining the good identification
of Glimepiride
(fig.4)
and Sitagliptin
(fig.1)
treated patients
(training,
validation
set patients and
parameter
lists)
fig.6 The most sensitive parameter fishing percentile pair
30/70% generates like the high accuracy pair 5/95% only HbA1c
as most discriminating parameter, indicating that Glimepiride by direct
action on the Langerhans cells of the pancreas influences only the
HbA1c amongst the 32 determined parameters of the trial
(training,
validation,
parameters).
fig.7 Like in the 5/95% classification
(fig.5) the median (7.30%) of the HbA1c
value distribution is right skewed as indicated by a mean of 7.43%.
4. Compressed Result Display
The use of confusion matrices
(figs
1/
2/
4)
is suitabe for the display of individual classifications but a compressed
display, containing the same information, is preferable.in case of multiple
classifications.
|
SAIS1 Clinical Trial: Quality of Individual Patient Classifications
|
drug |
classific |
perc (%) |
pat (nl/nv) |
spectp (%) |
senstp (%) |
npvtp (%) |
ppvtp (%) |
specvp (%) |
sensvp (%) |
npvvp (%) |
ppvvp (%) |
| Sita |
26/0wks |
5/95 |
38/35 |
97.4 |
62.9 |
74.0 |
95.6 |
100.0 |
75.0 |
83.3 |
100.0 |
| Sita |
26/0wks |
25/75 |
38/34 |
92.1 |
50.0 |
67.3 |
85.0 |
90.0 |
50.0 |
69.2 |
80.0 |
| Glime |
26/0wks |
5/95 |
43/38 |
95.3 |
65.8 |
75.9 |
92.6 |
83.3 |
54.6 |
66.7 |
75.0 |
| Glime |
26/0wks |
30/70 |
43/38 |
72.1 |
76.3 |
77.5 |
70.7 |
70.7 |
83.3 |
71.4 |
77.8 |
tab.1 The Sitagliptin (Sita) classifications
(fig.1) at the 5/95%
percentie levels is superior to the 25/75%
(fig.2).
classification, similarly as the 5/95% Glimepiride (Glime)
(fig.4)
to the 30/70%
(fig.6).
classification, concerning highest specificity, sensitivity,
negative and positive predictive values,indicating that the
highest quality classifications are obtained with the lowest
possible percentile pairs.
Abbreviations: Sita=Sitagliptin, Glime=Glimepiride,
classific=classification category, perc=percentile, pat=patient,
spectr, senstr, npvtr, ppvtp=specificity, sensitiity, negative/positive
predictive values of the training set, specva, sensva, npvva, ppvvp for
validation set patients.
5. Classifier Robustness: Cross Validation
CLASSIF1 classifiers are inherently robust as shown by cross-validating
(tab.4)
the embedded validation patients.
|
SAIS1 Clinical Trial: Cross-Validation of CLASSIF1 Classifications
(Sitagl.26 versus 0 weeks)
|
drug |
classific |
perc (%) |
pat (nl/nv) |
spectp (%) |
senstp (%) |
npvtp (%) |
ppvtp (%) |
specvp (%) |
sensvp (%) |
npvvp (%) |
ppvvp (%) |
| Sita |
p_56,60,65... |
5/95 |
38/35 |
97.4 |
62.9 |
74.0 |
95.6 |
100.0 |
75.0 |
83.3 |
100.0 |
| Sita |
p_57,61,66... |
5/95 |
38/33 |
97.4 |
66.7 |
70.1 |
95.7 |
100.0 |
60.0 |
71.4 |
100.0 |
| Sita |
p_58,62,67... |
5/95 |
38/35 |
97.4 |
62.9 |
74.0 |
95.7 |
100.0 |
75.0 |
83.3 |
100.0 |
| Sita |
p_59,63,68... |
5/95 |
38/34 |
97.4 |
64.7 |
75.5 |
95.7 |
100.0 |
66.7 |
76.9 |
100.0 |
| Sita |
p_60,64,69... |
10/90 |
38/34 |
97.4 |
64.7 |
75.5 |
95.7 |
100.0 |
66.7 |
76.9 |
100.0 |
tab.2 Validation patients are hidden to the learning process by
a question mark in the patient label record at the identifier position
(character
7).
The inaccessible character 8 of the 20 character patient label
contains the yellow labeled truth C=control or T=therapy.
for the learning process.
The hidden validation patients do in fact not appear in the training set
classification, as can be
verified
with patients nr. 56,60,65,70 .... missing
amongst the training set patients.
Validation patients were sequentially rotated, starting with
hidden patients 56,60,65,70... then 57,61,66,71 .. further 58,62,67,72 and so on,
followed by reclassification.
Classifier robustness is due to the elimination of all non informative parameters
during the learning process.
The last classification was performed with the 10/90% percentiles because
no classification was obtained at the 5/95% percentile thresholds.
6. Sitagliptin/Glimepiride prior Therapy (0weeks)
Patients groups in clinical studies are selected to match as good as
possible for a certain number of parameters like age, sex or previous
medical history but certain differences concerning investigaed molecular
parameters may still exist.
Sitagliptin patients were classified against Glimepiride
patients prior to therapy start (0weeks) to detect differences
between the patient groups.
|
CLASSIF1: SAIS1 Clinical Trial: Sitagliptin/Glimepiride
prior Therapy (0w)
|
| perc (%) |
pat (nl/nv) |
spectp (%) |
senstp (%) |
npvtp (%) |
ppvtp (%) |
specvp (%) |
sensvp (%) |
npvvp (%) |
ppvvp (%) |
| 10/90 |
43/38 |
86.1 |
44.7 |
63.8 |
73.9 |
75.0 |
40.0 |
60.0 |
57.1 |
| 30/70 |
43/38 |
88.4 |
50.0 |
66.7 |
79.2 |
83.3 |
40.0 |
58.8 |
60.0 |
tab.3 The
10/90%
classification does not show marked differences between both
groups because specificity, sensitivity and predictive values in the training
and validation groups of patients remain below the 90% accuracy level.
RLP
(remnant triglycerise lipoprotein) (p<0.05)
was picked as single parameter at the high discrimination
10/90%
percentile pair.
(training,
validation,
parameters).
Classification at the parameter picking
30/70%
percentile pair provided a
9 parameter classification mask -+++-+-+- for parameter differences
prior to therapy in the Sitagliptin patient group comprising
(-) FMD (flow-mediated dilation),
(+) BMI (body mass index),
(+) HBA1C (glycated hemoglobin A1c),
(+) PROINS (pro-insulin),
(-) TRIGL (triglycerides),
(+) TOTPAI-1 (total plasminogen activator inhibitor-1),
(-) NTPROBND (N terminal prohormone of brain natriuretic peptide),
(+) TNF-alpha (tumor necrosis factor-alpha).
(-) U-ALB (albuminuria)
(p<0.05,
<0.10,
>0.100)
when compared to the Glimepiride reference group classification mask:
000000000
(training,
validation,
parameters,
means,statistics).
7. Sitagliptin/Glimepiride after Therapy (26weeks)
|
CLASSIF1: SAIS1 Clinical Trial: Sitagliptin/Glimepiride
(26weeks)
|
| perc (%) |
pat (nl/nv) |
spectp (%) |
senstp (%) |
npvtp (%) |
ppvtp (%) |
specvp (%) |
sensvp (%) |
npvvp (%) |
ppvvp (%) |
| 20/80 |
38/34 |
81.6 |
52.9 |
66.0 |
72.0 |
63.6 |
42.9 |
63.6 |
42.9 |
tab.4 The lowest still classifying percentile pair
(20/80%)
provides
the single decreased (-) mask parameter
D-ROMS
(reactive oxygen metabolites-derived compounds)
as best classifying parameter
for the discrimination of Sitagliptin versus Glimepiride
effects at 26weeks.
(training,
validation,
parameters),
although the low quality classification of the validation patients (vp)
emphasizes the importance of determining further parameters.
|
CLASSIF1: SAIS1 Clinical Trial: Cross Validation
Sitagliptin/Glimepiride (26weeks)
|
perc
(%) |
pat (nl/nv) |
spectp (%) |
senstp (%) |
npvtp (%) |
ppvtp (%) |
specvp (%) |
sensvp (%) |
npvvp (%) |
ppvvp (%) |
| 25/75
val.pat: 1,5,10.. |
38/34 |
84.2 |
55.9 |
68.1 |
76.0 |
90.9 |
28.6 |
66.7 |
66.7 |
| 25/75
val.pat: 2,6,11.. |
39/32 |
89.7 |
53.1 |
70.0 |
81.0 |
90.0 |
10.1 |
52.8 |
50.0 |
| 25/75
val.pat: 3,7,12.. |
38/33 |
100.0 |
39.4 |
65.5 |
100.0 |
88.8 |
0.0 |
50.8 |
0.0 |
| 25/75
val.pat: 4,8,13.. |
40/32 |
85.0 |
46.9 |
66.7 |
71.9 |
77.8 |
33.3 |
53.8 |
60.0 |
| 25/75
val.pat: 5,9,14.. |
39/32 |
84.6 |
46.9 |
66.0 |
71.4 |
80.0 |
33.3 |
57.1 |
60.0 |
tab.5
Parameter fishing classifications at the
25/75%
percentile pair provided
(-) D-ROMS (reactive oxygen
metabolites-derived compounds) together with
(+) CRP (high sensitivity
C-reactive protein) as differences between Sitagliptin and
Glimepiride classifications,
using cross validation patients (valpats) 1,5,10...
(learning,
validation,
parameters,
means,statistics),
(-) CARDIX (cardiac output index:
total cardiac output (l/min)/body surface (m2)),
(+) TPRI (total
peripheral resistance index),
(-) D-ROMS
for valpats 2,6,11...
(+) TPRI,
(+) TNF_?,
(-) D-ROMS
for valpats 3,7,12...,
(+) TPRI,
(-) CARDIX
for valpats 4,8,13... and
(-) D-ROMS
for valpats 5,9,14...
(p<0.05,
<0.10).
Training and validation classifications are of low quality.
They indicate however in tendency a Sitagliptin associated
pattern of molecular differences in comparison to Glimepiride
action, consisting of reduced (-)
D-ROMS, increased proinflammation
(+) CRP, (+) TNF-alpha
(preexisting !)
as well as
(+) TPRI at decreased
(-) CARDIX
The parameters may be suitable indicators for further
studies on metabolic differences between the action of both
drugs.
8. Manual Classification
Standardized classifications
can be performed manually with known
percentile thresholds
of: <7.06, >1.66, >58.46, >5.83, >28.06 for the parameters selected by the
preceding classification process, like for example with the values
of patient 72 Hb1Ac=7.20, CPEPTIN=1.70, HDL=46.0, SOD=8.80, BAP=29.877
yielding the classification mask: 0+0++, classifying as "T" .
This is correct as can be verified in the
training
group (rec.51).
9. References:
1.
Nomoto1 H, Miyoshi H, Furumoto T et al
A Randomized Controlled Trial Comparing
the Effects of Sitagliptin and Glimepiride on
Endothelial Function and Metabolic
Parameters: Sapporo Athero-Incretin Study 1
(SAIS1).
PLoS ONE (2016) 11(10): e0164255. doi:10.1371/journal.pone.0164255
2.
Oyama J-i, Murohara T, Kitakaze M, Ishizu, Sato Y, Kitagawa K, et al.
The Effect of Sitagliptin on Carotid Artery Atherosclerosis in Type 2
Diabetes: The PROLOGUE Randomized Controlled Trial.
PLoS Med (2016) 13(6): e1002051. doi:10.1371/journal.pmed.1002051
3.
Berchialla1 P, Lanera C, Sciannameo V, Gregori D, Baldi T.
Prediction of treatment outcome in clinical trials under a personalized medicine
perspective.
Scient Rep (2022) 12:4115 | https://doi.org/10.1038/s41598-022-07801-4
4.
Valet.G
Human cytome project: A new potential for drug discovery.
In: Las Omicas genomica, proteomica, citomica y metabolomica:
modernas tecnologias para desarrollo de farmacos.
Ed: Real Academia Nacional de Farmacia, Madrid (2005) p 207-228
5.
Valet G, Valet M, Tschöpe D, Gabriel H, Rothe G,
Kellermann W, Kahle H.
White cell and thrombocyte disorders: Standardized, self-learning
flow cytometric list mode data classification with the CLASSIF1
program system.
Ann NY Acad Sci (1993) 677: 233-251
< back
Internet:
https://www.classimed.de/health1.html
gval22 last update: Jan 15,2025
first display: Dec 18,2024