Listen to Audio
Listen Anywhere
December 2025
In this episode, host Bob Chew shares insights on accelerating technology adoption in the pharmaceutical industry. Dive into this discussion to explore the future of pharma innovation, including transitioning from document-centric to data-driven quality systems, leveraging AI and digital twins for enhanced manufacturing processes, and reimagining risk management with statistical tools.
1
00:00:00,000 --> 00:00:05,000
Welcome to the ISPE podcast, shaping the future of pharma, where ISPE
2
00:00:05,000 --> 00:00:10,000
supports you on your journey, fueling innovation, sharing insights, thought leadership,
3
00:00:10,000 --> 00:00:14,000
and empowering a global community to reimagine what's possible.
4
00:00:15,000 --> 00:00:20,000
Hello, and welcome to the ISPE podcast, Shaping the Future of Pharma.
5
00:00:21,000 --> 00:00:22,000
I'm Bob Chew, your host.
6
00:00:23,000 --> 00:00:29,000
And today, we have another episode where we'll be sharing the latest insights
7
00:00:27,000 --> 00:00:32,000
and thought leadership on manufacturing, technology, supply chains
8
00:00:32,000 --> 00:00:36,000
and regulatory trends impacting the pharmaceutical industry.
9
00:00:37,000 --> 00:00:42,000
You will hear directly from the innovators, experts and professionals driving progress
10
00:00:42,000 --> 00:00:44,000
and shaping the future.
11
00:00:45,000 --> 00:00:49,000
Thank you again for joining us, and now let's dive into this episode.
12
00:00:50,000 --> 00:00:54,000
Our topic today is accelerating technology adoption.
13
00:00:55,000 --> 00:01:00,000
I am recording this podcast from a National Geographic ship in the Antarctic.
14
00:01:01,000 --> 00:01:06,000
If you hear occasional crashes or sounds like thunder, that is likely the
15
00:01:06,000 --> 00:01:08,000
ship moving through the ice pack.
16
00:01:09,000 --> 00:01:15,000
I have no special guest for this episode and will instead offer my thinking on this
17
00:01:15,000 --> 00:01:18,000
topic of accelerating technology adoption.
18
00:01:19,000 --> 00:01:24,000
As part of doing this podcast series, I attend many ISPE
19
00:01:24,000 --> 00:01:30,000
and other industry conferences, hearing case studies, listening to executives,
20
00:01:30,000 --> 00:01:35,000
regulators, and technical experts talk about the future and the present.
21
00:01:36,000 --> 00:01:41,000
I hope you will find this thought provoking and perhaps create your own innovative
22
00:01:41,000 --> 00:01:45,000
approaches to shape the future of pharmaceutical manufacturing.
23
00:01:47,000 --> 00:01:52,000
In this session, I'm going to discuss three broad topics: documents, data,
24
00:01:53,000 --> 00:01:54,000
and statistics.
25
00:01:55,000 --> 00:02:00,000
Now what do these have to do with the subject of this podcast, Accelerating Technology
26
00:02:00,000 --> 00:02:01,000
Adoption?
27
00:02:02,000 --> 00:02:08,000
Well, innovative technologies such as AI and digital
28
00:02:08,000 --> 00:02:13,000
twins are based on sophisticated statistical algorithms.
29
00:02:13,000 --> 00:02:19,000
When we train these models, we feed them data, lots of data, data
30
00:02:19,000 --> 00:02:22,000
that is relevant to what we want the model to do.
31
00:02:22,000 --> 00:02:27,000
The more data, the more statistically relevant outputs we will get.
32
00:02:28,000 --> 00:02:33,000
Of course, large language models are trained based on language, I e,
33
00:02:34,000 --> 00:02:35,000
documents.
34
00:02:36,000 --> 00:02:41,000
Now GMP regulations contain many references to
35
00:02:41,000 --> 00:02:45,000
and expectations of documents.
36
00:02:45,000 --> 00:02:49,000
Our quality systems are largely document based.
37
00:02:50,000 --> 00:02:55,000
We have policies, procedures, protocols, work instructions
38
00:02:56,000 --> 00:02:57,000
by the thousands.
39
00:02:59,000 --> 00:03:00,000
Let me ask this question.
40
00:03:01,000 --> 00:03:06,000
Should our quality systems continue to be document based or should we
41
00:03:06,000 --> 00:03:09,000
move towards a more data driven quality system?
42
00:03:11,000 --> 00:03:17,000
Now, our production and process control systems provide data
43
00:03:17,000 --> 00:03:20,000
by the boatload continuously.
44
00:03:21,000 --> 00:03:25,000
Separately, we have data generated by our QC laboratories.
45
00:03:26,000 --> 00:03:32,000
Separately, we have information about human performance, including training records.
46
00:03:33,000 --> 00:03:38,000
Today, technology exists to monitor operators and assess human performance,
47
00:03:39,000 --> 00:03:44,000
which could focus on how the operators contribute to process and contamination control
48
00:03:45,000 --> 00:03:50,000
and record proper or improper aseptic techniques, for example.
49
00:03:50,000 --> 00:03:56,000
Thus, we can even include information about human performance
50
00:03:56,000 --> 00:03:59,000
into the category of data.
51
00:04:01,000 --> 00:04:06,000
Now, new technologies and computing power exist to synthesize
52
00:04:06,000 --> 00:04:11,000
all the above and create a production model which could
53
00:04:11,000 --> 00:04:16,000
analyze the effectiveness of the process and contamination control strategies
54
00:04:17,000 --> 00:04:22,000
and even identify opportunities for improvement through statistical evaluation.
55
00:04:24,000 --> 00:04:29,000
In other words, data, a data driven quality system.
56
00:04:31,000 --> 00:04:36,000
Let's examine how data could supplant documents as the foundation of
57
00:04:36,000 --> 00:04:37,000
our quality system.
58
00:04:38,000 --> 00:04:43,000
Today, when we have a deviation, we create a document.
59
00:04:44,000 --> 00:04:49,000
That document describes what happened and offers root cause analysis,
60
00:04:50,000 --> 00:04:53,000
corrective actions, and preventive actions for the future.
61
00:04:54,000 --> 00:04:59,000
We agonize over the wording of these documents since they will be
62
00:04:59,000 --> 00:05:04,000
used to explain to future inspectors what really happened and
63
00:05:04,000 --> 00:05:09,000
provide a convincing case that the corrective and preventive actions were
64
00:05:09,000 --> 00:05:12,000
both justified and sufficient.
65
00:05:14,000 --> 00:05:20,000
What if in the future we had a combination of Process Digital Twin constantly
66
00:05:20,000 --> 00:05:25,000
monitoring the process and an AI platform that integrated
67
00:05:25,000 --> 00:05:31,000
past deviation data, past what happened, what
68
00:05:31,000 --> 00:05:35,000
were the corrective actions, and what were the results of those actions.
69
00:05:36,000 --> 00:05:42,000
When we have the next deviation, these technologies can provide a statistically
70
00:05:42,000 --> 00:05:47,000
ranked set of possible root causes with associated
71
00:05:47,000 --> 00:05:53,000
probabilities and recommended corrective actions that are based on observed
72
00:05:53,000 --> 00:05:56,000
effectiveness of similar actions in the past.
73
00:05:57,000 --> 00:06:02,000
The human can still provide the machine generated still approve the
74
00:06:02,000 --> 00:06:07,000
machine generated recommendation but using statistics based on data.
75
00:06:08,000 --> 00:06:14,000
And from time to time, the human may identify new
76
00:06:14,000 --> 00:06:19,000
root causes and new opportunities for innovative improvements.
77
00:06:20,000 --> 00:06:24,000
But we no longer agonize over the wording of documents.
78
00:06:27,000 --> 00:06:32,000
Further, today, our conclusions can often be operator error
79
00:06:32,000 --> 00:06:35,000
or laboratory error.
80
00:06:35,000 --> 00:06:41,000
If AI is assessing these situations based on data, it
81
00:06:41,000 --> 00:06:46,000
might point to other root causes based on the real data and the statistical
82
00:06:46,000 --> 00:06:51,000
probability that it really isn't operator error or
83
00:06:51,000 --> 00:06:53,000
laboratory error.
84
00:06:54,000 --> 00:06:59,000
Let's broaden the conversation and consider quality risk management.
85
00:07:00,000 --> 00:07:05,000
First, I'd like to remind ourselves that each and every GMP
86
00:07:05,000 --> 00:07:11,000
regulation is designed to control a risk to
87
00:07:11,000 --> 00:07:12,000
manufacturing quality.
88
00:07:13,000 --> 00:07:18,000
In other words, the original quality risk management system is our
89
00:07:18,000 --> 00:07:19,000
GMP regulations.
90
00:07:20,000 --> 00:07:25,000
For example, GMP regulations tell us that we must have procedures
91
00:07:25,000 --> 00:07:30,000
for cleaning and that those cleaning procedures
92
00:07:30,000 --> 00:07:32,000
be validated.
93
00:07:32,000 --> 00:07:34,000
Why is that?
94
00:07:34,000 --> 00:07:39,000
Well, it is to control the risk that cleaning is
95
00:07:39,000 --> 00:07:44,000
not sufficient and that it doesn't clean properly.
96
00:07:45,000 --> 00:07:48,000
And so we have these regulations.
97
00:07:49,000 --> 00:07:54,000
How might we use a statistically based technology to automatically
98
00:07:54,000 --> 00:07:56,000
generate risk assessments?
99
00:07:57,000 --> 00:08:02,000
Well, if we had an AI model of aseptic
100
00:08:02,000 --> 00:08:08,000
processing And that AI model
101
00:08:08,000 --> 00:08:13,000
was trained on a combination of regulations
102
00:08:13,000 --> 00:08:18,000
for aseptic process control like Annex one.
103
00:08:18,000 --> 00:08:23,000
If it was trained on how equipment is designed
104
00:08:23,000 --> 00:08:29,000
for aseptic, control, contamination control, if it was
105
00:08:29,000 --> 00:08:34,000
trained on the actual observed effectiveness of those controls
106
00:08:34,000 --> 00:08:39,000
during ongoing aseptic production operations, we
107
00:08:39,000 --> 00:08:45,000
could then have a tool that is automatically updated around
108
00:08:45,000 --> 00:08:48,000
the concept of contamination control.
109
00:08:51,000 --> 00:08:56,000
And it would be assessing risks, new risks,
110
00:08:56,000 --> 00:09:01,000
old risks, how effective are those controls, and constantly
111
00:09:01,000 --> 00:09:06,000
reporting out that these controls are effective and these controls
112
00:09:06,000 --> 00:09:09,000
might need, improvements.
113
00:09:11,000 --> 00:09:16,000
I think it would be super powerful if we could apply, statistical
114
00:09:16,000 --> 00:09:22,000
techniques through process digital twins, through aseptic
115
00:09:22,000 --> 00:09:27,000
process digital twins to generate automatically
116
00:09:27,000 --> 00:09:29,000
quality risk management metrics.
117
00:09:32,000 --> 00:09:37,000
Now how might equipment digital twins be used to accelerate startup
118
00:09:37,000 --> 00:09:39,000
and qualification of equipment?
119
00:09:40,000 --> 00:09:45,000
After all, digital twins of jet engines have been used for a long time to
120
00:09:45,000 --> 00:09:50,000
analyze performance and to model failure modes and effects.
121
00:09:51,000 --> 00:09:54,000
Well, imagine an aseptic filling line.
122
00:09:55,000 --> 00:10:01,000
Imagine a digital twin of such a filling line, which, by the way, has
123
00:10:01,000 --> 00:10:06,000
already been created and used in at least one aseptic manufacturing
124
00:10:06,000 --> 00:10:08,000
plant in Ireland.
125
00:10:08,000 --> 00:10:13,000
There was a case study presented at an ISPE conference about a year ago.
126
00:10:15,000 --> 00:10:19,000
Imagine training this digital twin during factory acceptance testing.
127
00:10:21,000 --> 00:10:26,000
And so now this digital twin has a representation,
128
00:10:26,000 --> 00:10:31,000
a model of how that aseptic line should work, how each and every
129
00:10:31,000 --> 00:10:37,000
manipulation on that line should work, its performance.
130
00:10:37,000 --> 00:10:43,000
Now we disassemble the line, we ship it, we set it up at the, manufacturing
131
00:10:43,000 --> 00:10:48,000
site, and as we go through the startup process, we plug
132
00:10:48,000 --> 00:10:50,000
in the digital twin.
133
00:10:51,000 --> 00:10:56,000
It monitors how the machine is operating, and it's able to point
134
00:10:56,000 --> 00:11:01,000
that you need to adjust, make this adjustment on this part.
135
00:11:01,000 --> 00:11:04,000
You need to look at this connection.
136
00:11:04,000 --> 00:11:05,000
It's not working right.
137
00:11:06,000 --> 00:11:11,000
But at the end of the day, it's able to automatically say
138
00:11:11,000 --> 00:11:14,000
that the equipment is qualified.
139
00:11:15,000 --> 00:11:16,000
Wouldn't that be sweet?
140
00:11:17,000 --> 00:11:23,000
Today, AI writes computer code and
141
00:11:23,000 --> 00:11:26,000
it self tests that computer code.
142
00:11:27,000 --> 00:11:32,000
Not necessarily in the pharmaceutical industry, but that technology exists.
143
00:11:32,000 --> 00:11:37,000
Imagine being able to apply that technology to automatically generate
144
00:11:37,000 --> 00:11:43,000
the process control automation software and
145
00:11:43,000 --> 00:11:45,000
to automatically test it.
146
00:11:47,000 --> 00:11:52,000
That would certainly be an acceleration of the delivery of manufacturing capability.
147
00:11:54,000 --> 00:11:59,000
Now I'm going to suggest a break from past approaches to process validation.
148
00:12:01,000 --> 00:12:07,000
First, unless you're doing a site to site transfer, it's
149
00:12:06,000 --> 00:12:12,000
unreasonable to expect that development batches supplemented
150
00:12:12,000 --> 00:12:17,000
with a few engineering batches will yield a statistically valid
151
00:12:17,000 --> 00:12:18,000
design space.
152
00:12:19,000 --> 00:12:20,000
Let me say that again.
153
00:12:21,000 --> 00:12:26,000
Design space is that three-dimensional or multi dimensional
154
00:12:26,000 --> 00:12:32,000
model of where you can manufacture and get a
155
00:12:32,000 --> 00:12:33,000
successful batch.
156
00:12:35,000 --> 00:12:39,000
That takes quite a few statistical, batches to build.
157
00:12:41,000 --> 00:12:46,000
Nor can we really say that a process is truly validated based on three successful
158
00:12:46,000 --> 00:12:52,000
batches, which are not always without a failed batch or two interspersed.
159
00:12:54,000 --> 00:12:59,000
I will offer that a robust process digital twin
160
00:12:59,000 --> 00:13:05,000
is nothing more than the design space from an underlying statistics
161
00:13:05,000 --> 00:13:06,000
perspective.
162
00:13:07,000 --> 00:13:11,000
It takes quite a few batches to train a digital twin.
163
00:13:12,000 --> 00:13:17,000
But once you have it trained, you also have a robust
164
00:13:17,000 --> 00:13:20,000
design space defined.
165
00:13:21,000 --> 00:13:26,000
As we then continue to manufacture and problems arise, the digital
166
00:13:26,000 --> 00:13:31,000
twin can either point to the root cause or
167
00:13:31,000 --> 00:13:36,000
it can say unknown new source of variation identified,
168
00:13:37,000 --> 00:13:42,000
which might be that you change bioreactor bag supplier or the supplier altered
169
00:13:42,000 --> 00:13:47,000
its methods or raw materials without understanding the impact on your process.
170
00:13:49,000 --> 00:13:54,000
In short, getting a process to a validated
171
00:13:54,000 --> 00:13:59,000
state is a continuous process and it will take
172
00:13:59,000 --> 00:14:05,000
quite a few batches to really develop that robust design space.
173
00:14:06,000 --> 00:14:11,000
Okay, you might agree with me theoretically, but from a practical perspective,
174
00:14:12,000 --> 00:14:16,000
when would regulators give the green light to market the product?
175
00:14:16,000 --> 00:14:19,000
How many batches must be produced first?
176
00:14:20,000 --> 00:14:25,000
Well, I can imagine a regulator having an AI tool
177
00:14:26,000 --> 00:14:31,000
that statistically assesses the degree of control exhibited by the
178
00:14:31,000 --> 00:14:37,000
process as reflected in consistent process performance data.
179
00:14:38,000 --> 00:14:43,000
The licensed applicant would have the same data and would notify regulators
180
00:14:43,000 --> 00:14:48,000
when it feels that it has its process in a full state of control, regardless
181
00:14:48,000 --> 00:14:50,000
of how many batches that is.
182
00:14:51,000 --> 00:14:56,000
The regulator would access the process data and,
183
00:14:56,000 --> 00:15:02,000
using its AI tool, analyze the data and render an approval decision.
184
00:15:03,000 --> 00:15:06,000
This would automate and accelerate the licensing process.
185
00:15:07,000 --> 00:15:12,000
Couple this with machine digital twins reporting out to regulators that a qualified
186
00:15:12,000 --> 00:15:17,000
state has been attained, and we take the guesswork and time delays
187
00:15:18,000 --> 00:15:20,000
out of, preapproval inspections.
188
00:15:24,000 --> 00:15:29,000
If we agree that new technologies based on statistics will drive us
189
00:15:29,000 --> 00:15:34,000
towards a data centric quality system, then data integrity
190
00:15:34,000 --> 00:15:38,000
takes on even more importance than it has up to now.
191
00:15:39,000 --> 00:15:44,000
Data integrity has focused on ensuring that data meets the ALCOA
192
00:15:44,000 --> 00:15:50,000
plus characteristics of attributable, legible, contemporaneous,
193
00:15:50,000 --> 00:15:56,000
original, accurate, complete, consistent, enduring,
194
00:15:56,000 --> 00:15:57,000
and available.
195
00:15:58,000 --> 00:16:04,000
But in addition, when it comes to training an AI or digital twin model, the
196
00:16:04,000 --> 00:16:09,000
data must also be relevant to the situation or application being
197
00:16:09,000 --> 00:16:10,000
trained.
198
00:16:10,000 --> 00:16:15,000
Now I will leave it to our Pharma four point o colleagues to expound upon this dimension
199
00:16:15,000 --> 00:16:18,000
of data and data integrity.
200
00:16:19,000 --> 00:16:24,000
So back to the title of this podcast episode, Accelerating Technology
201
00:16:24,000 --> 00:16:25,000
Adoption.
202
00:16:26,000 --> 00:16:28,000
How can we achieve this acceleration?
203
00:16:29,000 --> 00:16:34,000
My answer is, by recognizing how these new technologies really
204
00:16:34,000 --> 00:16:39,000
work, we accept the use of statistics to assess clinical outcomes,
205
00:16:39,000 --> 00:16:41,000
the basis for all drug approvals.
206
00:16:42,000 --> 00:16:48,000
We understand that quality risk management is founded on probabilistic estimates
207
00:16:48,000 --> 00:16:50,000
of hazards and controls.
208
00:16:51,000 --> 00:16:56,000
Therefore, we should embrace the power of new statistical tools of
209
00:16:56,000 --> 00:17:02,000
which AI and digital twins are prime examples and implement
210
00:17:02,000 --> 00:17:05,000
such tools across the manufacturing operation.
211
00:17:06,000 --> 00:17:12,000
In conjunction, we need to deemphasize documents and base
212
00:17:12,000 --> 00:17:14,000
our quality systems on data.
213
00:17:15,000 --> 00:17:20,000
Both industry and regulators should come together to discuss how
214
00:17:20,000 --> 00:17:25,000
to transform our industry to a nimble, data driven approach to process control,
215
00:17:26,000 --> 00:17:29,000
contamination control, and quality assurance.
216
00:17:30,000 --> 00:17:36,000
Old expectations of documents should be replaced by new expectations
217
00:17:36,000 --> 00:17:37,000
of data analysis.
218
00:17:39,000 --> 00:17:45,000
The quality function should evolve from one of compliance to procedures
219
00:17:45,000 --> 00:17:50,000
to one of a true quality engineer, driving innovation and changes for
220
00:17:50,000 --> 00:17:51,000
improvement.
221
00:17:52,000 --> 00:17:57,000
We must get comfortable with the use of AI and digital twins for
222
00:17:57,000 --> 00:18:02,000
automated process control and continued refinement of the design space.
223
00:18:04,000 --> 00:18:09,000
In summary, if we view AI and digital twins from the lens
224
00:18:09,000 --> 00:18:14,000
of statistics and if we can appreciate that the more data we feed these
225
00:18:14,000 --> 00:18:19,000
tools, the more useful and robust they become, we conclude
226
00:18:19,000 --> 00:18:24,000
that we should move away from a document centric quality system and towards a data
227
00:18:24,000 --> 00:18:26,000
centric quality system.
228
00:18:27,000 --> 00:18:32,000
Quality management throughout the life cycle of a product and facility becomes
229
00:18:32,000 --> 00:18:35,000
a continuous improvement journey.
230
00:18:36,000 --> 00:18:41,000
Each improvement, each change, should be made based on the statistics of
231
00:18:41,000 --> 00:18:47,000
real data and in most cases can be implemented either automatically
232
00:18:47,000 --> 00:18:49,000
or without a change control package.
233
00:18:51,000 --> 00:18:56,000
That brings us to the end of another episode of the ISPE podcast, Shaping
234
00:18:56,000 --> 00:18:58,000
the Future of Pharma.
235
00:18:59,000 --> 00:19:05,000
Please be sure to subscribe so you don't miss future conversations with the innovators,
236
00:19:05,000 --> 00:19:09,000
experts, and change makers driving our industry forward.
237
00:19:10,000 --> 00:19:15,000
On behalf of all of us at ISPE, thank you for listening, and we'll see
238
00:19:15,000 --> 00:19:20,000
you next time as we continue to explore the ideas, trends, and
239
00:19:20,000 --> 00:19:23,000
people shaping the future of pharma.
